Integrated Neuroscience and Neuroinformatics Approach for Understanding and Treating Brain Disorders Master’S Thesis

Home , Local field potential, Neuroinformatics, Neuroprosthetics

LITHUANIAN UNIVERSITY OF HEALTH SCIENCES MEDICAL ACADEMY FACULTY OF MEDICINE NEUROSCIENCE INSTITUTE LABORATORY OF BIOPHYSICS AND BIOINFORMATICS

Laura Sakalauskaitė

Integrated Neuroscience and Neuroinformatics Approach for Understanding and Treating Brain Disorders Master’s Thesis

Thesis Supervisor: Dr. Aušra Saudargienė

KAUNAS 2021

TABLE OF CONTENTS

1. SUMMARY ...... 3 2. ACKNOWLEDGEMENTS ...... 5 3. CONFLICTS OF INTEREST ...... 5 4. PERMISSION OF THE ETHICS COMMITTEE ...... 5 5. LIST OF ABBREVIATIONS ...... 6 6. INTRODUCTION ...... 8 7. AIM ...... 9 8. OBJECTIVES ...... 9 9. LITERATURE REVIEW ...... 10

9.1 FUNDAMENTAL METHODS OF NEUROINFORMATICS IN BRAIN RESEARCH ...... 10 9.1.1 Neuroscience Data Analysis...... 10 9.2 COMPUTATIONAL NEUROSCIENCE ...... 12 9.2.1 A Scientific Case for Brain Simulations ...... 12 9.2.2 Levels of Analysis ...... 13 9.3 TOWARDS PERSONALIZED MEDICINE ...... 18 9.3.1 Parkinson’s Disease and Deep Brain Stimulation ...... 18 9.3.2 Epilepsy ...... 22 9.3.3 Alzheimer’s Disease ...... 25 9.3.4 Stroke ...... 28 9.3.5 Computational Psychiatry ...... 29 9.3.6 Brain-machine Interfaces ...... 30 10. RESEARCH MATERIALS AND METHODS ...... 32 11. RESULTS AND DISCUSSION ...... 36 12. CONCLUSIONS ...... 43 13. PRACTICAL RECOMMENDATIONS ...... 44 14. REFERENCES ...... 45 15. ANNEXES ...... 56

1. SUMMARY

Author: Laura Sakalauskaitė Research Title: Integrated Neuroscience and Neuroinformatics approach for understanding and treating brain disorders. Aim: To review the neuroinformatic methods used in the clinical domain of brain sciences and investigate the impact of deep brain stimulation on pathological brain activity in Parkinson’s disease using a computational modeling approach. Objectives of Study: 1. To review the main concepts and methods in the field of neuroinformatics: aspects of neural data analysis and multiscale modeling in computational neuroscience. 2. To review the advances of the integrated neuroscience and neuroinformatics approach in understanding and treating complex neurological and psychiatric disorders. 3. To investigate the influence of different deep brain stimulation (DBS) targets and parameters on pathological brain oscillations in Parkinson’s disease by utilizing a computational modeling approach. Methodology: A literature review of machine learning methods and computational modeling approaches in the clinical domain of neuroscience was conducted. Furthermore, a computational modeling study was performed to investigate deep brain stimulation effects on pathological beta band oscillations of a simulated parkinsonian network using a neural mass model that represents basal ganglia and thalamocortical connections. Results: Machine learning and computational modeling applications in Parkinson ‘s disease, epilepsy, Alzheimer ‘s disease, stroke and psychiatric disorders were identified. A chosen computational model showed that connections to the thalamus and cortex were essential for driving pathological oscillations, increasing STN (subthalamic nucleus) and GPe (globus pallidus externus) connection strength led to higher frequency activity and applying DBS stimulation to selected targets with effective parameters drove parkinsonian oscillations to higher frequency bands. Conclusions: Computational and machine-learning approaches contribute to understanding the neuronal mechanisms and dynamical processes of brain diseases, offering individualized virtual brain models for early diagnosis, treatment selection and outcome prediction. The computational model of Parkinson’s disease revealed the principal network connections for driving pathological oscillatory activity. Modeling deep brain stimulation effects on the network allowed to investigate the effects of DBS parameters and multitarget stimulation. Recommendations: Multidisciplinary collaboration of neuroscience research fields, clinical medicine and neuroinformatics is needed for developing theoretical and computational models that integrate patient data and create individualized brain models that lead to novel clinical applications and hypothesis testing frameworks.

1. SANTRAUKA

Autorė: Laura Sakalauskaitė Darbo pavadinimas: Integruotų neuromokslų ir neuroinformatikos metodų taikymas tiriant ir gydant nervų sistemos ligas. Tikslas: Apžvelgti neuroinformatikos metodų taikymą klinikinėje neuromokslų srityje ir panaudojant kompiuterinį modelį ištirti giliosios smegenų stimuliacijos efektus Parkinsono ligos sukeltai patologinei smegenų veiklai. Uždaviniai: 1. Apžvelgti pagrindinius neuroinformatikos aspektus: duomenų analizės metodus ir daugiaskalinius modeliavimo lygmenis kompiuterinių neuromokslų srityje. 2. Apžvelgti integruotų neuromokslų ir neuroinformatikos metodų panaudojimo galimybes tiriant ir gydant nervų sistemos ligas ir psichikos sutrikimus. 3. Ištirti skirtingų giliosios smegenų stimuliacijos taikinių ir parametrų įtaką Parkinsono ligai būdingiems patologiniams smegenų signalams taikant kompiuterinio modeliavimo metodus. Metodologija: Apžvelgta literatūra apie kompiuterinio modeliavimo ir mašininio mokymo metodus klinikinėje neuromokslų srityje. Kompiuterinis modelis, aprašantis smegenų pamato branduolių, gumburo ir smegenų žievės neuroninius tinklus, buvo naudojamas tiriant giliosios smegenų stimuliacijos poveikį beta bangų dažnio pokyčiams Parkinsono ligos atveju. Rezultatai: Mašininio mokymo ir kompiuterinio modeliavimo metodai yra plačiai taikomi Parkinsono ligos, epilepsijos, Alzheimerio ligos, insulto ir psichikos sutrikimų tyrimuose. Pasirinktas Parkinsono ligos kompiuterinis modelis parodė, kad jungtys į gumburą ir smegenų žievę buvo pagrindiniai komponentai, skatinantys patologinius smegenų signalus, didėjantis STN (subtalaminio branduolio) ir GPe (išorinio globus pallidus) ryšio stiprumas lėmė didesnio dažnio bangų aktyvumą ir taikant GSS su efektyviais parametrais į pasirinktas zonas buvo panaikinti patologinio dažnio smegenų signalai. Išvados: Kompiuteriniai modeliai padeda suprasti sudėtingus ligų mechanizmus ir skatina naujų terapinių galimybių ir personalizuoto gydymo atsiradimą neurologijos, neurochirurgijos ir psichiatrijos srityse. Pasirinktas Parkinsono ligos neuroninio tinklo kompiuterinis modelis parodė, kurie neuroninio tinklo ryšiai skatino patologinę smegenų veiklą. Giliosios smegenų stimuliacijos modeliavimas leido ištirti optimalius stimuliacijos parametrus ir skirtingų smegenų branduolių stimuliacijos poveikį. Rekomendacijos: Norint sukurti teorinius ir kompiuterinius smegenų modelius, kurie integruoja individualius pacientų duomenis, ir pritaikyti šiuos modelius klinikinėje praktikoje paciento gydymo plano parinkimui bei hipotezių tikrinimui, būtinas tarpdisciplininis neuromokslinių tyrimų sričių, klinikinės medicinos ir neuroinformatikos specialistų bendradarbiavimas.

2. ACKNOWLEDGEMENTS

The completion of this thesis would not have been possible without the guidance and expertise of my thesis supervisor Dr. Aušra Saudargienė.

3. CONFLICTS OF INTEREST

There are no conflicts of interest.

4. PERMISSION OF THE ETHICS COMMITTEE

No clearance issued by the Ethics Committee was needed in this study.

5. LIST OF ABBREVIATIONS

AD - Alzheimer’s Disease AEDs - anti-epileptic drugs APP - amyloid precursor protein Aβ - amyloid beta CNN - convolutional neural network CR - coordinated reset CSF - cerebrospinal fluid Cx - cortex DCN - deep cerebellar nuclei DSM - Diagnostic and Statistical Manual of Mental Disorders ECoG - electrocorticography FDA - Food and Drug Administration GPe - globus pallidus externus GPi - globus pallidus internus GSK3 - Glycogen synthase kinase 3 HC - healthy controls HIP - Human Intracerebral EEG Platform ICD - International Classification of Disease iEEG - intracerebral electroencephalogram ILAE - International Legue Against Epilepsy k-NN - k-nearest neighbour LFP - local field potential MCI - mild cognitive impairment MEG - magnetoencephalography MRI - magnetic resonance imaging NFT - Neurofibrillary tangles NMDA - N-methyl-D-aspartate receptor nRT - thalamic reticular nucleus

PET - positron emission tomography PHFs - paired helical filaments PLV - phase locking value PZ - propagation zone RDoC - Research Domain Criteria RNN - recurrent neural network SOZ - seizure onset zone STN - subthalamic nucleus Th - thalamus TVB - The Virtual Brain VEP - Virtual Epileptic Patient Vim - ventral intermediate nucleus of thalamus WHO - World Health Organization

6. INTRODUCTION

In the last few decades there has been remarkable progress in deciphering the fundamental mechanisms of the human brain. The development of a complete color-coded map of the C. elegans nervous system containing 302 neurons and all their connections [1], the use of light-activated proteins to study the functionality of neuronal networks on multiple scales of living organisms in optogenetics research [2], light-sheet fluorescence expansion microscopy for visualising neural circuits at super high resolutions [3] or investigating the genetic basis of human brain disorders in an emerging field of neurogenetics [4] are just some examples. With neuroscience fields generating large empirical data sets, a demand in computational data analysis and storage tools resulted in the emerging field of neuroinformatics supported by advancement in high performance computing systems (HCP) [5].

Despite intensive neuroscience research and impressive technological advances, the burden of deaths and disability caused by neurological and psychiatric disorders is increasingly being recognised as a global public health challenge and with aging and growing populations the burden is expected to rise [6]. The current scarcity of established modifiable risks for most of the neurological burden demonstrates that new methods of knowledge integration are required to develop effective prevention and personalized treatment strategies. In the light of these issues, to synthesize the knowledge acquired through research and with the help of current advances in neuroimaging and acquisition and processing of very large data sets many worldwide neuroscience initiatives have emerged [7]. To outline some examples, the European Union’s Human Brain Project (HBP) has implemented multidimensional neuroinformatics databases and infrastructure platforms for large-scale biologically realistic simulations, neuromorphic engineering, robotics, and medical data integration tools [7]. US based Allen Institute for Brain Science has contributed a comprehensive, cellular-level atlas of gene expression in the adult laboratory mouse, neuronal cell type database based on single-cell transcriptional, morphological, electrophysiological, connectional, and functional properties [7, 8]. The China Brain Project and its “one body, two wings” research framework sets an emphasis on studying the neural basis of cognitive functions while improving and developing diagnoses and therapies for major brain disorders and brain- machine technologies. [7]. Finally, Japan Brain/MINDS focuses on structural and functional mapping and modeling neurodevelopmental and neurodegenerative diseases of Marmoset brain, developing translatable brain markers for neuropsychiatric disorders. Many other worldwide initiatives such as The Brain Initiative (USA), Australian Brain Alliance, Canadian Brain Research, Israel Brain Technologies, and Korea Brain project are participating in a worldwide collaboration, information and tool sharing to advance our understanding of the organisation and function of the brain, and potentially the treatment of mental and neurological disorders. 8

This study is a part of the project “Prediction of neurosurgical treatment outcomes in Parkinson’s disease” (The Human Brain Project’s EBRAINS Research Infrastructure Voucher Programme 2021- 2022) and has two major purposes: 1. To review the principles and examples of neuroinformatic methods and tools implemented in the clinical domain of brain sciences: advances in Neurology, Psychiatry and Neurotechnology fields. 2. To apply the methods described in the literature review and use a computational model of neuronal population activities to investigate the impacts of deep brain stimulation on pathological brain activity in Parkinson’s disease.

7. AIM

To review the neuroinformatics methods used in the clinical domain of brain sciences and investigate the impact of deep brain stimulation on pathological brain activity in Parkinson’s disease using a computational modelling approach.

8. OBJECTIVES

1. To review the main concepts and methods in the field of neuroinformatics: aspects of neural data analysis and multiscale modeling in computational neurosciences.

2. To review the advances of the integrated neuroscience and neuroinformatics approach in understanding and treating brain disorders: Parkinson’s disease, epilepsy, Alzheimer’s disease, stroke and psychiatric disorders.

3. To investigate the influence of deep brain stimulation (DBS) targets and parameters on pathological brain oscillations in Parkinson’s disease by utilizing the computational modeling approach.

9. LITERATURE REVIEW

9.1 Fundamental Methods of Neuroinformatics in Brain Research

With technological advancement, such as availability of very high-resolution imaging, large and multi-dimensional collections of neuroscience data are being generated every day [9]. To make use of this data, the process of sharing, storage and analysis becomes crucial. Neuroinformatics is an interdisciplinary field that combines methods from computer sciences, statistics, engineering, and integrative biology to analyse, understand and derive meaning from neuroscience data Fig 1. Furthermore, The International Neuroinformatics Coordinating Facility (INCF) develops and endorses the best standards and practices for data usage, sharing and reproducibility worldwide.

Fig. 1 The main pillars of Neuroinformatics

9.1.1 Neuroscience data analysis

Data analysis is an integral part of neuroscience research. For example, analysis of EEG recordings, high-dimensional brain imaging of multiple modalities (fMRI, PET, CT) or gene expression patterns in neurological and psychiatric disorders [5, 10]. To deal with the noisy and complex data sets, next to the traditional statistics, neuroscience has been utilizing machine learning methods such as artificial neural networks and more recently – deep neural networks [11, 12].

Machine learning is one of the subsets of artificial intelligence (AI) and is based on a computer or “machine” automatically learning and making predictions based on the “training” data without any explicit instructions or rules. These learning tasks are mainly classified into two categories: supervised

10 and unsupervised machine learning Fig. 2. Supervised learning uses labeled input and output training data for predicting the classification of the unlabeled data using machine learning algorithms. This ML method mainly consists of classification and regression analysis problems and algorithms used for predicting continuous or discrete values. Examples of popular algorithms utilized in supervised learning include linear regression which predicts a target value based on independent variables. For classification purposes, algorithms such as logistic regression, Naïve Bayes, support vector machine (SVM) or k- nearest neighbor are being used [13]. On the contrary, unsupervised learning is an approach where an algorithm learns patterns using unlabeled data. It is mainly used for finding new features in big data sets which are useful for further categorization. Main techniques include clustering, dimensionality reduction and association analysis [14]. Clustering algorithms such as k-means are used to group objects into clusters based on similarity, while dimensionality reduction and a method of principal component analysis is based on reducing the number of dimensions in data while retaining important information, finding patterns, and providing better visualization. Examples of this contain areas of face recognition and image compression [15]

Reinforcement learning is another application of ML and is based on agents taking actions in a particular environment to maximize the cumulative reward and it does not require a training data set to provide useful insights [11, 16]

Fig. 2 Classification of Machine Learning Methods with commonly used algorithms KNN - k-nearest neighbor classifier, SVM - support vector machine, SVD - singular value decomposition, PCA – principal component analysis, MLP - multilayer perceptron, CNN - convolutional neural network, RNN - recurrent neural network, LSTM - long short-term memory.

Artificial neural network (ANN) is a type of machine learning algorithm inspired by the human nervous system and attempts to mimic how biological neurons communicate and transmit signals. It is also employed in both supervised and unsupervised machine learning methods. Just like neuronal cells, artificial neurons or nodes receive input, make computations, and produce an output. A deep neural network (DNN) is a ML approach that involves an ANN composed of at least a few separate layers of artificial neurons and using these models is termed deep learning. Apart from conventional DNN models such as multi-layer perceptron Fig. 3, complex network architectures including convolutional neural networks (CNN) for image and signal analysis and recurrent neural networks (RNN) and long-short term memory (LSTM) for time series analysis are being extensively used for neuroscience data [17–19].

Fig. 3 Structural characteristics of a Multilayer Perceptron (MLP)

9.2 Computational neuroscience

Computational Neuroscience is an interdisciplinary field which employs mathematical models, theoretical concepts and abstractions of the brain to understand the principles that govern the development, structure, physiology, and cognitive abilities of the nervous system. It serves as a theoretical framework for investigating the function and mechanisms of the nervous system in healthy and pathological states [20].

9.2.1 A scientific case for brain simulations

A clear challenge of neuroscience advancement is its complex subject - the brain. To our current knowledge, human brain contains 86 billion neurons with an average of 100 trillion connections to other neurons called synapses. In other scientific fields that contain complex subjects, models are created to explain how things interact, connect, and how systems could look like in the future. For instance,

12 multiscale weather simulations bridge scales from tens of meters to the size of our entire planet to accurately predict the weather [21]. This is where computational neuroscience is being utilized when researching the human nervous system with its complex interconnections. Through digital reconstructions and simulations, researchers can conduct in silico experiments, test, and generate many hypotheses and theories, improve experimental methods, make predictions, and provide principles for neuromorphic systems and neurorobotics which in turn could lead to more efficient large-scale simulations [22–24]. The integrated multilevel approach of simulating the brain in detail could lead to new breakthroughs in treating and diagnosing complex brain disorders.

9.2.2 Levels of analysis

Activities in our brain are multiscale in space and time, ranging from nanometres to meters and microseconds to hundreds of years. For this reason, possible simulations take place at several separate organization levels in the brain, ranging from a single molecule up to the whole brain level Fig. 4. To increase understanding of disease mechanisms at multiple levels scientists are utilizing a “bottom-up” approach which begins with biological knowledge about cells and circuits to determine and explain how these mechanisms support mental and behavioural phenomena [25]. In contrary, top-down brain modeling approach is based on theories derived from information gathered from cognitive data or brain recordings (e.g., fMRI, EEG, single cell recordings) and infers hypotheses about underlying mechanisms. Mathematical models exist to explain or describe the workings of the nervous system at all these levels [26].

Fig. 4 Levels of analysis

The Hodgkin–Huxley model

The basis for developing neuronal models, more specifically neuronal electrical activity, starts with a simple description of the plasma membrane at its biophysical properties. A semi-permeable

13 neuronal membrane that functions to maintain different concentrations of ions inside and outside the cell is modeled as an electrical circuit. As a traditional circuit, the model contains a capacitor, resistor, and a battery. The membrane is represented as a capacitor separating the charges on either side, ion channels are modelled as resistors and the tendency of ions to move back and forth is described by the Nernst potential as the electrical circuit’s battery. The Hodgkin–Huxley (HH) model Fig. 5 with an electrical circuit, Nernst potential and conductances as voltage and time dependent parameters has been widely used to represent the basis of neuronal function - the action potential and has become a standard technique in making detailed and biophysically realistic computational models [27, 28].

Fig. 5 Hodgkin-Huxley Model circuit representation of the neuronal membrane. Taken from Isakovic et al, 2018 [28].

Compartmental model of neurons

A complex morphology of neurons can be captured using multi-compartmental approach. Such models are descriptive of neuron’s biophysical properties and considered biologically accurate. Sections of neurite are represented by simple geometric shapes such as cylinders which are constructed of interconnected electrical circuit models Fig. 6 [29]. For example, neuronal models with an electrical field being introduced to cortical cells of different morphologies were represented by multi- compartmental models to study the precise effects of transcranial electrical stimulation (tES). The method investigated exact mechanisms of how electrical stimulation targets the cells at threshold and sub-threshold levels which could provide insights into more targeted use of stimulation in neurological and psychiatric disorders [30] .

Fig. 6 Morphologically detailed models of neurons (a) Reconstructed morphology acquired from stained neurons. (b) Cylindrical compartments are representing in the cable model of a neuron (c) Electrical circuits are used to describe each compartment. Adapted from Stefanou et al., 2016 [29].

Integrate-and-fire model

Simple models, such as Integrate and Fire, represent an action potential (spike) which is generated when the membrane potential reaches a certain threshold without detailed descriptions of the mechanisms behind it. Simplified models are useful in integrating into and studying large networks, since they are computationally more efficient and include only two differential equations. The Allen Institute for Brain Science has produced a comprehensive, publicly available database of predictive neuron models, such as Leaky integrate and fire (GLIF) and more detailed biophysical models for large simulations with a future perspective to model neurological and psychiatric disorders and understand how elements give ruse to behaviour, perception, and consciousness [31].

Intracellular mechanisms and synapses

A great number of modeling studies, at varying levels of detail, have aimed to explain synaptic plasticity which underlies the complex mechanisms of learning and memory. Less detailed phenomenological models attempt to explain the mechanisms behind learning and memory processes [32], while biophysical models contain biological details and attempt to accurately represent the intracellular and synaptic mechanisms of plasticity. For example, models of pathways leading to

15 phosphorylation of AMPAR subunits found in hippocampal pyramidal cell spines, which results in an increase in synaptic strength or long-term potentiation (LTP) through an increase in receptor channel conductance. This increase is counteracted by competing pathways that dephosphorylate the subunits, resulting in long term depression (LTD) [33]. Furthermore, models include calcium concentration and activation of calcium-dependent protein kinases and phosphatases which results in the induction of LTP/LTD. As calcium influx through the n-methyl-d-aspartate (NMDA) receptor plays a fundamental role in the induction of LTP/LTD, changes in the properties of NMDA receptor-mediated calcium influx will dramatically affect activity-dependent synaptic plasticity [34].

Neuronal populations and neural mass models

These models represent the activity of large populations of neurons and synapses and are being utilized in computational sciences for analysing brain rhythms. They have successfully been used to fit neuroimaging data, analyse EEG rhythms, epileptic brain dynamics, deep brain stimulation and are a major building block in the Virtual Brain [35]. For instance, The Epileptor is a phenomenological neural mass model (NMM) able to reproduce temporal dynamic of a seizure. It consists of 5 linear differential equations with five state variables describing different kinds of electrical discharges [36].

Another example, a Jansen-Rit neural mass model representing three interconnected neural populations: pyramidal projection neurons, excitatory and inhibitory interneurons forming feedback loops [37]. The research by Stefanovski et al., showed the implementation of this NMM in Alzheimer’s disease, by modeling impairment of inhibitory neuronal subpopulations caused by Abeta and anti- NMDAergic effects. The model was also capable of simulating brain activity recorded by local field potentials, EEG and other modalities [38].

Lastly, Wilson-Cowan model describes the dynamics of interactions between populations of very simple excitatory and inhibitory model neurons. It portrays overall activity of a large-scale neuronal networks, using just two main differential equations. Important model parameters consist of strength of connectivity between each subtype of population (excitatory and inhibitory) and the strength of input to each subpopulation. Varying these generates insight into various dynamical behaviors that are representative of observed activity in the brain such as multistability, oscillations, traveling waves, and spatial patterns [39]. In my thesis, I attempt to use this approach to model deep brain stimulation on the basal ganglia and thalamocortical networks affected in Parkinson’s disease.

Networks

Networks consisting of nodes or units are connected and associated with weights (w) which depend on the relative importance to the other connected inputs. Examples of neural networks range from less detailed models of associative memory to multi-compartmental models of the basal ganglia and their use in understanding Parkinson’s disease, effects of deep brain stimulation and transcranial magnetic stimulation [30, 40, 41]. Recently, whole-brain networks are being developed using functional and diffusion weighted MRI scans for reconstructing individual structural and functional brain network connections. Brain is parcellated into different regions constituting network nodes and network connectivity strength is derived from tractography of white matter fibres measured with DWI (diffusion weighted imaging). The dynamics of brain activity are simulated by using mathematical models that sufficiently describe the underlying neuronal interactions Fig 7.

Fig. 7 A schematic diagram of the functional and structural network analysis. Adapted from Lee et al., 2015 [6].

Computational neuroscience encompasses not only models described above, but effective data sharing, visualisation and analysis platforms, the use of artificial intelligence and knowledge learned from biological systems for increasing computational efficiency. To establish effective clinical 17 applications and gain thorough understanding of pathological processes scientists aided by neuroinformatic tools are attempting to connect these levels of analysis by multiscale multilevel modeling approaches.

9.3 Towards personalized medicine

The complexity of linkages that produces pathophysiology in neurological and psychiatric diseases requires a multiscale approach of investigation. Before reviewing the novel applications towards personalized treatments, it is important to address the current limitations in medicine. The prevalence of neurological diseases such as stroke, epilepsy and neurodegenerative diseases are high [6, 42] . The exact mechanisms of pathology and how it affects our brain on the network level is not known. Treatments are based on managing or slowing down symptoms, but not treating the underlying causes [6]. Psychiatric disorders are classified based on the symptoms, which is proven to produce inaccurate diagnoses and incorrect medication use [43]. Drug adherence issues, polypharmacotherapy with severe side effects and the increasing need of invasive techniques present the need for integrated brain research approach [44, 45]. Computational neuroscience methods and collaborative multiscale research has made progress in linking the mechanisms underlying the disease and has provided insights into more targeted treatment approaches that will be described in the following section.

9.3.1 Parkinson’s Disease and Deep Brain Stimulation

Parkinson's disease (PD) is a neurodegenerative disorder that affects dopaminergic neurons of a specific brain region called substantia nigra. The basal ganglia, a group of interconnected subcortical nuclei responsible for voluntary motor control, action selection, learning and cognitive tasks, heavily depend on the neurotransmitter dopamine. The interconnections between the nuclei, including the direct, indirect and hyper-direct pathways are represented in Fig. 8. The prevalence of this disease is increasing, and more than 10 million people are living with PD worldwide [46]. The cardinal symptoms are resting tremor, bradykinesia, rigor, and postural instability. In addition to PD effects on the motor system, a great variety of non-motor manifestations, such as sleep disturbances and cognitive impairment have been identified and related to complex brain network changes following the dopamine depletion in the brain [47]. L-3,4-Dioxyphenylalanine (levodopa) has been the gold standard for the treatment of Parkinson's Disease, although the effectivity of the drug over time decreases, the incidence of motor fluctuations and side effects such as dyskinesias, confusion and other life-limiting symptoms increases [48]. 18

Computational models have investigated pathophysiological mechanisms of PD from many perspectives [41, 49–51]. One example, dopamine depletion in these models disrupts the balance of the pathways leading to action selection deficits. The imbalance is hypothesized to arise from neural excitability or through aberrant corticostriatal plasticity that follows dopamine depletion [52, 53]. Not only dopamine depletion is investigated with computational models, but the mechanisms of dopaminergic therapy and its side effects. Frank et al., have shown that the medication impaired patient’s ability to learn from negative decision outcomes which would explain the incidences of pathological gambling activities [54].

Fig. 8 Basal ganglia and thalamocortical tracts and their subdivision into direct, indirect and hyperdirect BG pathways. Glutamatergic (excitatory) connections are marked by arrows, GABAergic (inhibitory) effects are represented with circles. Taken from Schroll et al, 2013 [41].

In the past the initial surgical management of PD was limited to non-reversable lesioning procedures (thalamotomy, pallidotomy), but in the last 25 years, deep brain stimulation has become the predominant therapy [55]. Deep brain stimulation (DBS) is a neuromodulatory intervention that has had a profound impact on movement disorders, including parkinsonism, essential tremor, and dystonia. Clinical trials experiment on utilising DBS for disorders such as obsessive–compulsive disorder, epilepsy, and even treatment-refractory depression [56–59]. Although the basic principle of DBS is to modulate pathological neural activity with applied electric fields, the mechanism by which DBS improves disease symptoms is still poorly understood and therapeutic parameters are still largely derived by trial-and-error [60]. Research has shown probable mechanisms of affecting information transmission between brain structures, disrupting pathologic oscillations, and inducing long-term plasticity [61–63].

One example, Creed and colleagues reversed cocaine-induced plasticity in the nucleus accumbens (NAc) of rodents with DBS [64]. They found that DBS successfully suppressed sensitization responses caused by repeated exposure to cocaine. These findings demonstrate the potential of DBS to effect and induce neuroplasticity and structural alterations of neural networks.

In terms of PD the usual DBS targets are the subthalamic nucleus (STN) and globus pallidus internus (GPi). STN is usually a choice when the goal of the surgery is to reduce dopaminergic medication and targeting GPi is recommended for reducing the severity of “on” medication dyskinesias and in case of cognitive decline or a high risk of depression [65]. Other targets such as ventral intermediate nucleus (Vim), and pedunculopontine nucleus (PPN) are reportedly effective DBS targets for control of Parkinsonian disease. Vim is an effective target when the dominating symptom is tremor, and low-frequency stimulation of PPN has proven to be helpful for the axial symptoms of PD (e.g., postural instability, gait dysfunction) [66–68]. Computational models of DBS including detailed biophysical individual neuron models to study the effects of impulses on the cells surrounding electrodes [69, 70], and less detailed network models for investigating STN, GPe and GPi connections and transmission of bursting activity and the disruption of the information flow through the thalamus have brought insights into the mechanisms of stimulation [71–73].

New developments in DBS devices have influenced an increase of the degrees of freedom for DBS programming and a surge of computational studies aiming to optimize them. Companies such as Medtronic, Boston Scientific and Abbott are continuously updating the DBS technologies, for example, octipolar multiple-independent current control stimulation design (Versice), multidirectional leads with current steering possibilities (“Infinity” by St Jude’s Medical Ltd.) or 40 configurable contact SureStim lead (Medtronic) [74–76]. These advances allow sculpting the field of stimulation, conforming to complex shapes, and precisely targeting substructures, e.g., motor components of STN. Furthermore, Percept Neurostimulator technology Fig. 9 allows to capture and record brain signals (local field potentials) while simultaneously delivering therapy to patients [70]. Optimization of DBS technology, from an engineering and clinical perspectives, will require improved scientific understanding of the effects and therapeutic mechanisms of electrical stimulation of the brain. Computational modeling is playing an important role in new developments to improve both electrode placement and stimulation parameter selection in patients undergoing DBS. Effects of different pulse widths, wave forms, amplitudes, frequencies, electrical field directions and other applicable parameters could be tested in a computational environment and hypotheses about network level effects could bring invaluable information into clinical management and device programming tailored to individual patients. For example, lower pulse widths were found to increase the therapeutic window and minimize the unwanted

20 side effects [69, 73, 77], multitarget stimulation was shown to require lower amplitudes and battery consumption [66, 69, 78].

Fig. 9 LFP recordings in Percept device Taken from reference source [79].

One novel DBS programming paradigm explored in theoretical studies called Coordinated Reset (CR) is based on an idea to specifically target the pathological beta band oscillations Fig. 10. The technique elucidated long-lasting therapeutic effects in parkinsonian monkeys as well as in patients with Parkinson’s disease in a clinical trial [79, 80]. Directional CR DBS combines a new DBS lead technology and a strategy of reproducing a comparable—or even better—therapeutic effect than traditional DBS therapies and with fewer side effects. Because it uses a very low stimulation intensity and significantly reduces battery consumption. Furthermore, closed loop approaches, where subject specific responses to stimulation are utilized to optimize stimulus parameters and patterns [81, 82].

Fig. 10 Coordinated reset (CR) neuromodulation technique Note: In CR neuromodulation, brief high-frequency pulse trains are delivered through different sites at different, equally spaced times. The 3:2 ON-OFF pattern2,4 and the random variation of the pulse train sequences5 optimize the desynchronizing CR effect. The pattern is repeated periodically. Taken from Adamchic et al., 2014 [80].

A very recent study [83] used multiscale modeling with The Virtual Brain platform (TVB) to tailor deep brain stimulation for individual patients. A co-simulation with spiking network model of the basal ganglia and their interactions with the cortical regions based on published normative connectivity tractography atlas [84] was implemented to demonstrate biologically plausible dynamics of PD and the effects of deep brain stimulation. Study showed changes in thalamic activity and altered cortical activity predominantly in frontal regions after the stimulation. This multiscale co-simulation approach has the potential for more advanced and personalized computational simulations and optimization of DBS lead placement and parameter configurations.

In this thesis I use a network model of the basal ganglia and thalamocortical networks to demonstrate how a computational model provides a framework to test different parameter settings and gain a mechanistic understanding of the effects of local DBS stimulus to the entire network of neurons.

9.3.2 Epilepsy

Epilepsy is characterized by spontaneously recurring seizures, which are a transient occurrence of signs and/or symptoms due to abnormal excessive or synchronous neuronal activity in the brain [85]. Around 50 million people worldwide have epilepsy, making it one of the most common neurological diseases globally [86]. Methods to treat epilepsy include medication, brain stimulation, surgery, dietary therapy, or various combinations of the above, directed toward the primary goal of eliminating or suppressing seizures [87]. Although around 30 different medications are developed to prevent seizures and alleviate the symptoms, approximately one third of the patients have refractory epilepsy (or so-called intractable seizures) that does not respond to anti-epileptic drugs (AEDs) [42, 88, 89]. The most common 22 solution for these patients is surgical removal of the area where seizure emerges called the “epileptogenic zone” with success rates of being seizure free post-surgery (based on Engel surgical outcome scale) at around 60% [90, 91].

Advances in technology in the past 20 years have dramatically influenced the translational research in epilepsy. Electroencephalography (EEG), high-resolution imaging and other patient data are now recorded digitally in most centres. Furthermore, free, and open-source platforms for sharing and using various types of data have been established, for example Human Intracerebral EEG Platform (HIP), dedicated to human intracerebral EEG (iEEG) and OpenNeuro for MRI, MEG, EEG, iEEG, ECoG, and ASL modalities [92, 93]. Following these developments, machine-learning approaches have been utilized for more accurate diagnosis, prediction, and detection of seizures and seizure onset zones.

Determining seizure and epilepsy types of individual patients is a crucial step for selecting the appropriate treatment. Studies have implemented several machine learning approaches including SVM, k-NN, deep learning algorithms such as CNN that automatically execute those diagnostic tasks and evaluated their performance [94–96]. For example, convolutional neural networks (CNN) were used to classify seizure types into 8 categories with accuracy up to 84.06% based solely on patient EEG recordings [95].

Seizure detection and prediction using machine-learning algorithms are important aspects of investigation especially in the context of developing on-demand preventive interventions. Seizure detection is based on identifying the seizure onset and seizure prediction forecasts the occurrence of seizure activity before the onset by recognizing differences of pre-ictal and inter-ictal states. In a recent study by Daoud and Bayoumi a CNN and RNN based model was able to predict seizures 1 hour before the onset, with a high accuracy of 99.6%, using long-term scalp EEG [97].

When a surgical intervention is considered for the treatment of seizures, an important step of investigation is detection of seizure onset and propagation zones (SOZ, PZ). In clinical practice these zones are usually identified using intracranial or stereotactic EEGs (iEEG). To aid this process and considering varied outcomes of surgery many ML approaches have been developed. One research based on features extracted from the phase locking value (PLV), effectively identified electrodes within the SOZ [98]. Therefore, ML approaches have the potential to assist clinicians in surgical decision-making when pre-surgical intracranial recordings are utilized.

The possibility to detect seizures prior to their onset creates a possibility for on demand management. Currently, research concentrates on effects of direct closed loop stimulation [100, 101]. On the contrary to open-loop systems where therapy is delivered according to pre- programmed settings and is unaffected

23 by changes in the patient’s symptoms or underlying disease processes, closed-loop neurostimulation systems modulate therapy in response to physiological changes usually detected by ML algorithms and may provide more effective and efficient therapy. The two neurostimulation methods approved by FDA are closed-loop responsive cortical stimulation and Vagus nerve stimulation, although deep brain stimulation (DBS) has been confirmed as a possible treatment option for refractory epilepsy in EU [99– 102]. The exact mechanisms of how these treatments work are unknown and integrative collaborative computational approaches are crucial for defining the best parameters of these modalities for maximum effectivity considering inherent patient-to-patient variability. A study by Sandler et al., used a functional closed-loop model of the hippocampus for investigating how physiological and epileptic oscillations emerge and designed efficient neurostimulation patterns to abate such oscillations [103]. The model was used to identify the optimal frequency parameters. This example shows that data-driven computational models have the potential to be utilized as a testbed for designing optimal DBS patterns for individual patients in these newly emerging treatment techniques.

A distinct branch of computational neuroscience uses more detailed biophysical models to describe underlying mechanisms in epilepsy. An important breakthrough in studying the pathophysiology of seizures was made with the help of nonlinear dynamical systems theory to identify the causes of seizure activity [26, 104]. Epilepsy is viewed as a dynamical process and mathematical theories, including multistability and bifurcations aid in explaining the brain’s transition into a seizure state. Based on these concepts, Jirsa and colleagues created a taxonomy of distinct types of seizures according to their onset and offset characteristics [105, 106].

Furthermore, mathematical, and pathophysiological knowledge was combined with patient data in the possibility of applications in surgery. Using a large-scale simulation engine, The Virtual Brain (TVB) personalized brain models are derived from non-invasive imaging techniques with individually measured anatomy and connectivity of each patient. Data is fitted and translated into a dynamical patient specific network model where a clinician can test different hypotheses and simulate resection of seizure onset zones Fig. 11. In 2019, a clinical trial EPINOV started using the Virtual Epileptic Patient (VEP) technology in surgery planning with promising initial results [35, 36, 107].

Fig. 11 Representation of the Virtual Epileptic Patient (VEP) modeling (a) Connectivity derived from DTI, largescale network model. (b) Cortical and subcortical surfaces. (c). Brain areas in model with increased epileptogenicity. (d) Exemplary simulated times series of symptomatic/ asymptomatic seizure. Taken from Falcon et al, 2016) [35].

Computational neuroscience methods for epilepsy research span from constructing a network based on microscale units including single neurons, synapses and ion channels to a whole-brain network model based on macroscale brain connectome to understand its functions and disease mechanisms. Unifying these levels of investigation with personalized patient data is a crucial step towards effective clinical applications [24]

9.3.3 Alzheimer’s disease

Worldwide, around 50 million people have dementia, and there are nearly 10 million new cases every year. Alzheimer's disease is the most common form of dementia and may contribute to 60–70% of cases [108]. While the prevalence of this disease increases, our knowledge of underlying mechanisms is still insufficient to provide early prognostic tools and targeted treatments [109]. Current management of AD consists of classification according to disease severity and symptomatic treatment [109]. Working memory and long-term declarative memory are affected early during the course of the disease. As the disease progresses, a person will develop severe memory impairments, neuropsychiatric symptoms and lose the ability to carry out everyday tasks. Considering the effects on quality of life, hight mortality

25 rates and limited treatment options many efforts in neuroscience research have been made to explain the causes and symptoms of the disease.

Alzheimer's disease is a heterogenous multifactorial disorder with multiple phenotypes and underlying mechanisms, therefore it is difficult to understand how the interactions of all possible mechanisms lead to the pathogenesis of the disease and to make conclusions about possible effective treatments based on isolated approaches. The example of interacting disease mechanisms and possible therapeutic approaches is represented in Fig. 12. Mathematical and computational models are of great value in providing a unifying framework of AD research by connecting information from various levels of detail, ranging from biochemical to whole-brain models.

Classical ML methods such as support vector machine (SVM) have attempted to provide imaging biomarkers for AD patients. By grouping subjects into healthy controls, AD patients and MCI (mild cognitive impairments) groups, ML methods have successfully managed to detect significant differences and correctly classify healthy vs AD groups, but not MCI [110]. Furthermore, data collected from patients could be utilized for even more accurate disease prediction and trajectory purposes. Data sets spanning from genes, proteins, clinical records and various imaging modalities could be combined in order to provide personalized mechanistic disease signatures for individual patient. Platforms such as EBRAINS and NeuroMMSig are providing invaluable tools for this type of data sharing, analysis, and integration into large-scale modeling [111] .

Many hypotheses exist to explain Alzheimer’s disease pathology that leads to degeneration of neurons and disruption of synapses within cortical and subcortical areas. Amyloid plaques and neurofibrillary tangle (NFT) accumulations have been considered the governing mechanisms of the disease and frequently referred to as the Amyloid and Tau hypotheses [112]. To briefly summarize, the amyloid cascade hypothesis postulates that the main mechanism of AD is when the transmembrane molecule amyloid precursor protein (APP) is cleaved by an enzyme called beta-secretase leaving insoluble monomer tangles which bind to form amyloid plaques. The hypothesis was supported by the discovery that the mutation in APP gene in chromosome 21 which leads to early-onset familial type of AD. The other theory concentrates on a microtubule protein Tau hyperphosphorylation which results in microtubule dissociation and formation of paired helical filaments (PHFs) and NFTs. The homeostatic mechanisms within the brain are complex and interconnected, therefore many other mechanisms of pathology have been discovered, including neuroinflammation, angiopathy, oxidative stress, overactive GSK3 and others [113, 114]. Computational models of AD exist to investigate the kinetics, mechanistic pathways and fibrillogenesis of Aβ, mechanisms of plaque formation, the kinetics of APP processing, GSK-3b, tau-based modeling, immunity-based modeling with anti- Aβ antibodies and etc [115].

Furthermore, utilizing computer-Aided Drug Design of β-Secretase, γ-Secretase and Anti-Tau Inhibitors and approaches to designing multi-target drugs with structure-based design, data mining/repurposing, and in silico screening for the Discovery of Novel Alzheimer's Therapeutics and disease modifying effects [116].

Fig. 12 Scales of AD pathology and possible treatment targets Taken from Stefanovski et al., 2021 [117].

In consideration of possible clinical applications, it is important to note the main manifestations of Alzheimer’s disease – memory and other cognitive deficits. According to some studies, plaques are found in the brains of many elderly people with normal cognition [118], therefore it is important to investigate how plaques in AD patients lead to cognitive symptomatic disease course. One large-scale modeling research attempts to explain the effects of altered molecular pathways on the neuronal populations and network dynamics. With the help of The Virtual Brain (TVB) platform the amyloid beta plaques measured by positron emission tomography (PET) are connected to a neural mass model and address the phenomena of hyperexcitability observed in AD. The principle aspects of modeling are represented in Fig. 13. Results demonstrated that the virtual AD group showed slower frequencies in simulated local field potentials and EEG compared to mild cognitive impairment (MCI) and healthy control (HC) groups. In addition to slowing phenomena, modeling the N-methyl-D-aspartate (NMDA) receptor antagonism of memantine in local population models, revealed potential functional reversibility of the observed large-scale alterations represented by the EEG slowing in virtually simulated in silico brains of AD patients. [35, 38, 117].

Fig. 13 Model of Abeta accumulation effects in Alzheimer’s disease Molecular Abeta pathway in AD cause hyperexcitation and disinhibition in the neural mass model. Taken from Stefanovski et al., 2019 [38].

9.3.4 Stroke

Stroke is a neurological event of a sudden onset due to primary problems in the vasculature. Strokes still rank as the third-highest cause of death in the western world [119]. These neurological events are largely responsible for mild to severe functional and behavioural deficits, often resulting in life-long treatment and rehabilitation, of which mechanisms and effectivity is not understood very well.

Computational models of the neural processes such as synaptic plasticity and learning systems are an essential part of researching stroke and gaining mechanistic understanding of post-stroke recovery. Models have described probable mechanisms allowing to maintain and recover function such as compensatory contribution of homologous regions in the contralateral hemisphere, increased interhemispheric connectivity and interactions between differentially specialized networks [120, 121].

One large-scale example is The Human Brain project’s attempt to utilize TVB for modeling stroke. Models of stroke patient brains showed parameter changes associated with poor long-term motor recovery. These included local hyper-excitability, lower conduction velocities in cortico-cortical connectivity, and an imbalance between local and global dynamics. Following the discovery of these parameters, The Virtual therapy is used to determine the potential therapeutic value of restoring the affected biophysical parameters (global coupling, conduction velocity, coupling between inhibitory and excitatory populations). The model of an individual patient’s brain can act as an indicator of pathogenic processes and responses to therapeutic intervention. The integration of this modeling to new technological advances in biomedical imaging can yield unprecedented information on brain structure

28 and function, providing a promising application for neuroprotective therapies—strategies to reduce the damage immediately after a stroke occurs [122, 123].

Computational stroke models aid in explaining the structural-functional connections and how focal damage could lead to a variety of symptoms and changes in the entire brain network. Furthermore, models of recovery could provide paradigms for neuroplasticity after stroke, probing rapid network reorganization of motor and language functions via neuromodulation and neuroimaging, or prognosis and virtual therapy with changing parameters that could aid in a more effective and individual neurorehabilitation.

9.3.5 Computational psychiatry

Heterogeneity is a key feature of all psychiatric disorders that manifests on many levels, including symptoms, disease course, underlying biology with a high dependency on the surrounding environment. Main diagnostic tools in current psychiatry practice consist of evaluating a patient's symptoms and classifying based on DSM or ICD manuals [43]. Research showed that psychiatric diseases are more complex, and this approach led to mis-diagnosing patients, drug adherence issues, increase of refractory disorder cases and the need of very invasive management techniques such as electroconvulsive therapy [44, 44, 45].

Studies in many domains have discovered important insights into mechanisms behind psychiatric disorders and possible effective treatments. A big step towards improving our understanding and management in the field of psychiatry - Research Domain Criteria project by US National Institute of Mental Health. RDoC is a research framework for investigating mental disorders. It integrates many levels of information (from genomics and circuits to behaviour and self-report) to explore basic dimensions of functioning that span the full range of human behaviour from normal to abnormal [124].

In terms of computational psychiatry, the usual division consists of data-driven and theory-driven approaches. Data-driven approaches consist of applying machine-learning techniques to high- dimensionality data to improve classification of disease, predict outcomes or even improve treatment selection. Theory-driven approaches concentrate on generating and testing hypotheses about the underlying mechanisms of psychiatric disorders and possible targeted management [125, 126].

Data-driven approaches have made an influence on diagnostic classification, prediction of treatment responses and optimal treatment selection. For example, MRI data was used to distinguish schizophrenia patients from healthy controls by using machine – learning algorithms. The best entries reached an area

29 under the curve (AUC) for classification of validation data of 0,89 [40, 127]. This could result in possible diagnostic imaging biomarkers for early diagnosis of schizophrenia. Another example in treatment selection was a study that extracted features from patient’s EEG in order to predict the most successful medication for the treatment of depression. This automatic algorithm outperformed the clinical evaluation approach [128].

Computational insight into mechanisms span from very detailed accurate neuronal network models to abstract networks and concepts. For example, using the existing knowledge of neurobiological changes in schizophrenia, a computational model of cortical pyramidal neurons and GABAergic interneurons demonstrated the effects of reducing NMDA receptor density on inhibitory interneurons which led to disinhibition and working-memory disturbances [129, 130]. Furthermore, other models investigate NMDA and GABA synaptic dysfunctions to explain the impairment in gamma frequency oscillations in schizophrenia patients [131].

9.3.6 Brain-machine interfaces

Brain-machine interface (BMI) is a technology that allows direct communication between the brain and an external device. The principles of BMI technology are visually represented and described in Fig. 14. The earliest example of this type of device is considered to be a cochlear implant. By processing and transmitting signals generated by sound this device restores hearing and has been successfully used in patients with sensorineural hearing loss [132]. The industry of brain-machine interface technologies is emerging and advancing with many possible applications in treatment and neuroprosthetics. One impressive example, an epidural wireless brain machine interface was used in a tetraplegic patient with C4-C5 level spinal cord injury [133]. Two epidural recorders with 64 electrodes bilaterally were implanted over the sensorimotor areas of the brain responsible for controlling movements. The electrocorticography (ECoG) signals were processed with an adaptive decoding algorithm, which is a set of mathematical models that deal with non-stationery signals and recalibration. After the algorithm processes the signals, commands are sent to the exoskeleton device and allow patient to move. For two years of the trial, the patient was able to cortically control the program that simulated walking and other limb movements. Furthermore, a neurotechnology company launched in 2016 called Neuralink has made impressive technological improvements in electrode design and placement surgery. The thread design contains thousands of electrodes (traditional leads of DBS devices contain 4 electrodes) and minimally invasive robotic surgery done with local anaesthesia is used to place them in selected areas of the brain.

In addition, animal studies are showing promise in the ability to accurately record neuronal spikes from detailed areas of the brain and interface with a computer to execute tasks by using the mind. This exemplifies the future role of brain machine technology in treating patients with a wide range of disabilities and disorders [134].

Fig. 14 Principles of BMI technology

A biological signal measured with EEG, ECoG, MUA is recorded and decoded by the system which predicts or estimates some abstract aspect of a person’s cognitive state. The result is translated into an action of the output device. Subjects can often observe the effects and modulate their brain signals to accomplish the desired tasks. MUA – multiunit activity. Adapted from Edelman et al., 2015 [135].

10. RESEARCH MATERIALS AND METHODS

A computational modeling approach of neural networks implicated in Parkinson’s disease was utilized to investigate the deep brain stimulation (DBS) effects on pathological brain oscillations. After reviewing the literature, a computational model was chosen to represent the populations involved in thalamocortical and basal ganglia networks [71].

Although biophysical models and detailed networks consisting of compartmental neurons are being used to research PD and DBS [49, 62, 69, 83, 136], these models are complex and require high computational resources., therefore a less detailed modeling approach representing big populations of neurons was chosen to investigate the networks of the Parkinsonian brain in this study [39, 71]. The selected model meets the criteria and objectives identified in the thesis in the following aspects:

1. The model shows moderate complexity, adequately describes the large-scale dynamics researched in the thesis and is computationally efficient which allows for integration with individual patient data and clinical applications. 2. The model is described using the neural mass formalism where equations capture the mean dynamics of large-scale temporal activity relevant in PD symptoms and treatment. 3. Dynamics of the basal ganglia and thalamocortical networks are highly dependent upon the interaction of inhibitory and excitatory cells. The chosen Cowan-Willson approach [137] consists of two variable descriptions for both types of populations. Furthermore, it includes the representation of external stimuli (e.g., deep brain stimulation). 4. The model is able to reproduce the clinically observed neuronal population behavior. It has been proven useful to describe empirical electrophysiological recordings including local field potentials (LFPs).

The thalamocortical basal ganglia network in the model consists of 4 excitatory and 3 inhibitory populations. Excitatory populations include the cortex, deep cerebellar nuclei (DCN), subthalamic nucleus (STN) and ventral intermediate nucleus (Vim) of the thalamus. Inhibitory populations in the model consist of internal and external parts of the globus pallidus (GPi and GPe) and the reticular nucleus (nRT) of the thalamic population. It is important to note that the DCN population is modeled only to project to the thalamus (Vim population), receives external input but is not dependent on the dynamics of other network populations Fig. 15. To describe the basic interactions between the inhibitory and exhibitory populations included in the chosen network the Wilson-Cowan approach [39, 138] is implemented. The neurons within one population are considered to be in close proximity and only temporal dynamics are represented in the selected approach. 32

Fig. 15 Representation of thalamocortical basal ganglia network Black arrows represent excitatory connections, round arrowheads – inhibitory connections. The red arrows indicate the four brain regions targeted with deep brain stimulation (DBS). The external drive is coming to DCN. nRT – reticular nucleus of thalamus, STN – subthalamic nucleus, GPe – globus pallidus externus, GPi – globus pallidus internus, DCN – deep cerebellar nuclei. Adapted from (Yousif et al, 2020). [71].

Seven differential equations are used to define the populations in our neural mass model:

푑퐸퐶푥 휏 = −퐸 + (푘 − 퐸 ) ⋅ 푍 (푤 퐸 ) 퐶푥 푑푡 퐶푥 푒 퐶푥 푒 1 푇ℎ

푑퐸푇ℎ 휏 = −퐸 + (푘 − 퐸 ) ⋅ 푍 (푤 퐸 − 푤 퐼 + 푤 퐸 − 푤 퐼 ) 푇ℎ 푑푡 푇ℎ 푒 푇ℎ 푒 2 퐶푥 3 푛푅푇 4 퐷퐶푁 5 퐺푃𝑖

푑퐼푛푅푇 휏 = −퐼 + (푘 − 퐼 ) ⋅ 푍 (푤 퐸 ) 푛푅푇 푑푡 푛푅푇 𝑖 푛푅푇 𝑖 6 퐶푥

푑퐸퐷퐶푁 휏 = −퐸 + (푘 − 퐸 ) ⋅ 푍 푒푥푡 퐷퐶푁 푑푡 퐷퐶푁 푒 퐷퐶푁 푒

푑퐼퐺푃푒 휏 = −퐼 + (푘 − 퐼 ) ⋅ 푍 (푤 퐸 − 푤 퐼 ) 퐺푃푒 푑푡 퐺푃푒 𝑖 퐺푃푒 𝑖 7 푆푇푁 8 퐺푃푒

푑퐼퐺푃𝑖 휏 = −퐼 + (푘 − 퐼 ) ⋅ 푍 (푤 퐸 ) 퐺푃𝑖 푑푡 퐺푃𝑖 𝑖 퐺푃𝑖 𝑖 9 푆푇푁

푑퐸푆푇푁 휏 = −퐸 + (푘 − 퐸 ) ⋅ 푍 (푤 퐸 − 푤 퐼 ) 푆푇푁 푑푡 푆푇푁 푒 푆푇푁 푒 10 퐶푥 11 퐺푃푒

*Ei - a proportion of firing neurons in an excitatory population at a point in time. Ij – a proportion of firing neurons in an inhibitory population at a point in time. τ – membrane time-constant of a population, set to 10ms. k -the maximum values of the response functions: ke = 0.9945 and ki = 0.9994. Zi and Ze – sigmoid firing rate functions.

The number of firing neurons in a population at a certain point in time are represented by Ei (i = Cx, Th, DCN or STN) and Ij (j = nRT, GPe or GPi). Ei describes the excitatory populations, Ij – inhibitory. Connection strength between two populations is represented by weights (wn) and is the average number of contacts per neuronal cell and the average current induced in the postsynaptic cell by a presynaptic action potential. The parameter was changed in the interval 0-30 to observe the influence of connection strength to the oscillatory dynamics. Ze (x) and Zi (x) are monotonically increasing sigmoid functions representing the proportion of cells firing in a population for a selected number of average membrane potential activity x(t).

1 1 푍푝(푥) = − 1 + 푒푥푝⁡(−푏푝(푥 − 휃푝)) 1 + 푒푥푝⁡(푏푝휃푝)

*p represents e or i, bp and θp are constants, and x – level of input activity. Wilson and Cowan approach parameters: θe = 1.3, be = 4, θi = 2.0, and bi = 3.7. The maximum values of the response functions: ke = 0.9945 and ki = 0.9994.

Table 1. Connection weights between neuronal populations

Note: Adapted from Yousif at al. (2020)

DBS input is given by a simple square pulse in the following equation:

1,001 4 1 퐷퐵푆(푡) = 퐴 ∑ 푠푖푛⁡(2푛휋푓푡), 휋 푛 푛=1,3,5

* A – amplitude in arbitrary units (a.u.), f – frequency, t – time. Square wave is the sum of sin waves with parameters above for the odd values of n.

An example equation showing DBS being applied to the STN population:

푑퐸푆푇푁 휏 = −퐸 + (푘 − 퐸 ) ⋅ 푍 (푤 퐸 − 푤 퐼 + 푫푩푺) 푆푇푁 푑푡 푆푇푁 푒 푆푇푁 푒 10 푆푇푁 11 퐺푃푒

The more detailed population characteristics were extracted using Fast Fourier transform (FFT) algorithm. The method allows to visualize the dominating frequency of a population at a given moment in time.

The model of thalamocortical basal ganglia network was implemented in the Python environment. The network was analysed under the following conditions:

1. Healthy network: thalamocortical basal ganglia network in control conditions without DBS. Parameters are presented in Table 1, Healthy state parameters column.

2. Parkinsonian network: parkinsonian network conditions without DBS input. The weights w4 (DNC → Thalamus) and w7 (STN → GPe) were reduced to 20 and 5 to generate parkinsonian oscillatory activity. Parameters are presented in Table 1, Beta band parameters column.

3. Restoring STN → GPe connection strength in a parkinsonian network: the weight parameter w7 (STN→ GPe) was increased from 5 to 20 while w4 (DCN→ Thalamus) connection was left unchanged. Weight parameters w4, w7 are presented in Table 1 and network connections in Fig. 15.

4. Parkinsonian network with DBS input to the STN population. DBS square pulses with the threshold parameters of 4 a.u. 50 Hz are applied to demonstrate change in the parkinsonian network dynamics.

5. Model representation of bursting activity with DBS input. Very low frequency pulses of 4 Hz are applied to the STN population in a parkinsonian brain network. Amplitude is unchanged (4 a.u.).

6. Simultaneous DBS input to STN and GPi populations of a parkinsonian network representing a multitarget stimulation approach. DBS square pulses with the effective threshold parameters of 3 a.u. and 50 Hz are applied to both populations at the same point in time to demonstrate changes in beta band activity.

11. RESULTS AND DISCUSSION

The thalamocortical basal ganglia network was used to investigate the influence of the DBS parameters and location on the oscillatory activity of cortical, STN, GPe, GPi neuronal populations in health and PD.

In Parkinson's disease electrophysiological changes of the basal ganglia and thalamocortical networks include changes in firing rates, increasing incidence of bursting behaviours, interneuronal synchrony and a beta-band oscillatory activity in a range of 13-35Hz [139, 140]. The gamma frequency band is defined as approximately 35 Hz-100 Hz, with the 40 Hz point being of particular significance for cognitive functions, memory, and emotion processing [141].

First, we analysed the dynamics of the healthy network without DBS input. Fig. 16, a)-d) shows cortical, GPi and STN populations with the dominating oscillation frequency of 44 Hz. The Parkinsonian model on the left, subplots e)-h) represent the beta band oscillations with cortical, GPi and STN population activity within the 20 Hz frequency range. Cortical, STN and GPi regions were selected because of their relevance in PD. GPi and STN are known to exhibit the most beta band oscillatory activity and are targeted in DBS treatment for PD [65, 73, 142].

By changing connection weights (wn) between different populations Table 1, the model could successfully transition from the healthy to the Parkinsonian state. Model showed that ascending network connections (External input → DCN → Thalamus → Cortex) are necessary for the beta band activity to persist and the reduction of STN → GPe connection strength results in a lower frequency activity. The ascending connection between the deep cerebellar nuclei and thalamus (Vim) was considered in the network, as DCN projects glutamatergic neurons to the thalamus resulting in hyperpolarization and inhibition of thalamic-motor system which leads to low frequency oscillations [67, 72, 143]. Cerebellar activity is a part of pathophysiology in relation to dysfunctional basal ganglia networks. The cerebellar interaction to the BG network is more complex and some studies have even suggested a compensatory role, since better motor and cognitive functions correlated with increased cerebellar connectivity [143, 144]. Another connection strength crucial for transitioning the network into beta band activity was STN → GPe. Theoretical and computational studies have demonstrated the importance of this connection architecture and its proneness to generate low frequency oscillations [69, 145]. Therefore, the weight parameter was changed in isolation to demonstrate the single connection’s influence on the entire network activity.

Fig. 16 Healthy vs Parkinsonian Brain Oscillations Healthy (a-d) and Parkinsonian (e-h) oscillations generated by the neural network model without DBS. Cortical, STN and GPi populations are shown in more detailed activity graphs. Frequencies of 44 Hz in a healthy and 20 Hz of a Parkinsonian model were observed in populations of the entire network. The connection weights between DCN → Thalamus and STN → GPe are decreased to generate the parkinsonian network activity. STN – subthalamic nucleus, GPi – globus pallidus internus, DCN – deep cerebellar nuclei, GPe – globus pallidus externus.

I demonstrated the changes in oscillatory activity when the weight parameter of STN → GPe (w7) connection was increased Fig. 17. The oscillatory activity increased in frequency from 20 Hz to around a normal range of 40 Hz. This reciprocal inhibitory – excitatory connection is a central part of basal ganglia and is widely researched in PD pathophysiology. Studies show the importance of abnormal firing rates and patterns on this network activity to be highly linked to motor dysfunction and changes in action selection processes [136].

Fig. 17 The effects of STN → GPe connection strength on the Parkinsonian network dynamics

The cortical, GPi, STN and GPe populations (a-d) show a change from beta band frequency of 20 Hz to higher frequency of 40 Hz. The activity graphs illustrate the importance of STN → GPe connection. The connection weight between STN and GPe is increased to 20 (from the original value 5). STN – subthalamic nucleus, GPe – globus pallidus externus, GPi – globus pallidus internus.

Furthermore, DBS via a square pulse, as described in the methodology chapter, was applied to the Parkinsonian network with beta band activity. As in the original article [71] the threshold amplitude parameter for changing low-frequency oscillations into high-frequency gamma range oscillations was 4 arbitrary units. An example of this change with more detailed population characteristics was demonstrated in Fig. 18, where 4 a.u. 50 Hz pulse was given to the STN population.

Fig. 18 DBS square pulses applied to the STN population

DBS stimulation to STN demonstrates changes in pathological network activity with a threshold amplitude of 4 a.u. With selected parameters of 4 a.u. 50 Hz (a), cortex, STN and GPi populations (b- d) transform to higher-frequency 50 Hz oscillations. STN – subthalamic nucleus, GPi – globus pallidus internus.

In clinical practice standard stimulation parameters range from 2 to 4 V amplitude, 30–450 µs pulse width and 130–185 Hz frequency [76, 146]. The network model is not representative of real clinical parameters but allows the exploration of parameter changes and their effects on the network dynamics. Not considering the therapeutic parameters, low frequency DBS pulses were applied to the parkinsonian network. The model was able to demonstrate a bursting activity pattern when DBS stimulation of very low frequencies (4 Hz) was applied Fig. 19. This correlates to the adverse reactions when patients with PD are stimulated with lower frequency pulses [147].

Fig. 19 Model representation of bursting activity

DBS with the parameters of 5 a.u. and 4 Hz (a). Low frequencies of stimulation did not change the amplitude of network activity, reduced the frequencies, and resulted in occurrence of bursting activity throughout the network. Bursting activity in the cortex, STN and GPi is shown in plots (b-d). STN – subthalamic nucleus, GPi – globus pallidus internus.

DBS generally uses high frequency stimulation to target a single nucleus, hence causing physical damage to the neural tissue. It also requires battery replacement because of high energy consumption and frequent side effects because of high-frequency stimulation to the adjacent zones. Some computational studies have explored various optimization protocols. Parameters, such as different waveforms and lower pulse widths have been proven to play an important role in optimizing the DBS effects [69, 77, 148]. Multi-target stimulation has also been investigated. Yu et al. proposed DBS stimulation targeted to two nuclei by injecting two kinds of low-frequency pulse stimulation currents with phase difference by placing two electrodes in the BG structures [69, 73, 78].

Lastly, by using the neural mass model, an attempt was made to compare network activity using two target (STN and GPi) stimulation. Results showed that double target stimulation managed to supress beta band activity at lower threshold amplitudes compared to single target (STN) approach. Fig. 20

Fig. 20 Multitarget stimulation

DBS targeted to STN (a-d) and to SNT and GPi simultaneously (e-h). Stimulation of the STN alone was not effective in supressing the pathological beta band activity of 20 Hz in cortical, STN and GPi populations. Simultaneous stimulation of STN and GPi regions leads to the change of low-frequency high amplitude activity to higher-frequency lower-amplitude activity in the range of 50 Hz. DBS parameters: 3 a.u. (arbitrary units) and 50 Hz.

When discussing the link between results generated by the model and experimental studies, DBS- like input influences the suppression of low frequency beta band activity and could be represented in a patient’s local field potential (LFP) recordings before and after stimulation [149]. The recorded population activity should also corelate to symptoms specific to Parkinson’s disease considering their proven correlation with dominating synchrony in the beta band frequency. A very promising match between the experimental data and simulation results has been obtained by the Human Brain Project team employing The Virtual Brain platform, and our current project is aiming to achieve and validate these results in the coming years.

To summarize, the neural mass model was able to demonstrate that network connections to the thalamus and cortex were the principal components for driving pathologic oscillatory activity and increasing the STN and GPe connection resulted in higher frequency oscillations. Modeling deep brain stimulation effects on the network allowed to investigate different parameters and multitarget stimulation approaches. PD and other movement disorders involve complex interactions between multiple circuits within the basal ganglia network, therefore targeting multiple circuits depending on individual symptoms might be an effective novel approach. Continued detailed research into cellular mechanism underlying DBS effects and improvements in electrode implantation and placement are crucial components of these studies.

12. CONCLUSIONS

Neuroinformatics is a field encompassing many aspects of neuroscience research, from data analysis tools and data sharing platforms to theoretical and computational models of neural systems and neurotechnology development. Machine learning approaches such as artificial and deep neural networks are widely used to analyse neural data, guide clinicians in decision making. Computational models of brain diseases bring far reaching insights into complex disease mechanisms, possible early diagnostic biomarkers, and more personalized treatment approaches. Worldwide initiatives, such as The Human Brain Project perfectly demonstrate the breakthrough possibilities of multilevel integrated brain research approach. It is executed by building virtual brain models based on individually measured structural and functional brain connections and combining it with representative mathematical models. This creates a functional platform to test hypothesis and select patient oriented decisions.

Multiscale modeling studies of epilepsy have provided tools for clinicians to automatically detect and classify seizure types, model seizure onset and propagation zones and perform targeted excision surgeries. Models of complex hypotheses in Alzheimer’s disease have aided in novel drug design, biomarker discovery and explaining pathological hyperexcitability mechanisms of beta amyloid accumulation leading to cognitive and working memory deficits. Computational studies involving stroke have provided insights into personalized rehabilitation approaches and more accurate outcome prediction based on changes in local and global parameters. In computational psychiatry, both data and theory driven approaches have provided more understanding of underlying mechanisms and lastly, brain-machine interface technologies have provided novel tools for prosthesis and neurotechnology.

Computational models have aided in our understanding of the mechanisms underlying deep brain stimulation and have helped drive new therapeutic approaches and more personalized treatment options in Parkinson’s disease. A chosen neural mass model was able to demonstrate the dynamics of the basal ganglia and thalamocortical networks in different states (Healthy and Parkinsonian) and important connections for driving pathological oscillatory activity. DBS effects and the influence of changes in parameter settings (amplitude, frequency) served as a valuable tool for investigating the mechanism of deep brain stimulation and effects on important brain networks. New advances in DBS technology are promising and will allow closed-loop neuromodulation therapies, which are able to adapt the stimulation based on real-time, symptom-specific, and task-dependent input signals. Hence, computational models will serve as integral tools for planning and optimizing novel and advanced neuromodulation therapies.

13. PRACTICAL RECOMMENDATIONS

Theoretical and computational analysis of neural networks is a useful tool for investigating the dynamical processes of neuronal population activities and the effects of deep brain stimulation in Parkinson’s disease. Computational models including indirect, direct, and hyper direct pathways suitable for the integration of the individual patient data with structural and functional connectivity and electrode placement artefacts, provide a powerful tool for effective DBS parameter selection and outcome prediction in Parkinson’s disease.

Multiscale biophysically realistic models could provide a more detailed and accurate understanding of Parkinson’s disease and deep brain stimulation mechanisms on a molecular, cellular, microcircuit, brain network and behavioural scales.

Considering the advancement of DBS devices and the free parameter space, exploration of different wave forms, direction of electrical fields, lower pulse widths and their effects on brain networks should be studied in more detail. Furthermore, when having devices able to record local field potentials (LFPs) and register symptoms at the same time, clinical effects could be more closely correlated to the signal changes. Even more personalized adjustment could be done contributing to the emerging closed loop techniques.

Development of useful and clinically applicable computational models requires a multidisciplinary collaboration between neuroscientists, engineers, and clinicians. Having background and knowledge of these concepts in neuroscience is essential and could drive substantial scientific breakthroughs.

14. REFERENCES

1. Cook SJ, Jarrell TA, Brittin CA, et al (2019) Whole-animal connectomes of both Caenorhabditis elegans sexes. Nature 571:63–71 2. Montagni E, Resta F, Mascaro ALA, Pavone FS (2019) Optogenetics in Brain Research: From a Strategy to Investigate Physiological Function to a Therapeutic Tool. Photonics 6:92 3. Bürgers J, Pavlova I, Rodriguez-Gatica JE, Henneberger C, Oeller M, Ruland JA, Siebrasse JP, Kubitscheck U, Schwarz MK (2019) Light-sheet fluorescence expansion microscopy: fast mapping of neural circuits at super resolution. Neurophotonics 6:1 4. Zoghbi HY, Warren ST (2010) Neurogenetics: Advancing the “Next-Generation” of Brain Research. Neuron 68:165–173 5. Kovatch P, Gai L, Cho HM, Fluder E, Jiang D (2020) Optimizing High-Performance Computing Systems for Biomedical Workloads. In: 2020 IEEE Int. Parallel Distrib. Process. Symp. Workshop IPDPSW. IEEE, New Orleans, LA, USA, pp 183–192 6. Deuschl G, Beghi E, Fazekas F, Varga T, Christoforidi KA, Sipido E, Bassetti CL, Vos T, Feigin VL (2020) The burden of neurological diseases in Europe: an analysis for the Global Burden of Disease Study 2017. Lancet Public Health 5:e551–e567 7. Grillner S, Ip N, Koch C, Koroshetz W, Okano H, Polachek M, Poo M, Sejnowski TJ (2016) Worldwide initiatives to advance brain research. Nat Neurosci 19:1118–1122 8. Stockton DB, Santamaria F (2017) Integrating the Allen Brain Institute Cell Types Database into Automated Neuroscience Workflow. Neuroinformatics 15:333–342 9. Landhuis E (2017) Neuroscience: Big brain, big data. Nature 541:559–561 10. Auger SD, Jacobs BM, Dobson R, Marshall CR, Noyce AJ (2021) Big data, machine learning and artificial intelligence: a neurologist’s guide. Pract Neurol 21:4–11 11. Botvinick M, Wang JX, Dabney W, Miller KJ, Kurth-Nelson Z (2020) Deep Reinforcement Learning and Its Neuroscientific Implications. Neuron 107:603–616 12. Cichy RM, Kaiser D (2019) Deep Neural Networks as Scientific Models. Trends Cogn Sci 23:305– 317 13. Uddin S, Khan A, Hossain ME, Moni MA (2019) Comparing different supervised machine learning algorithms for disease prediction. BMC Med Inform Decis Mak 19:281 14. Sidey-Gibbons JAM, Sidey-Gibbons CJ (2019) Machine learning in medicine: a practical introduction. BMC Med Res Methodol. https://doi.org/10.1186/s12874-019-0681-4

15. Raza K, Singh NK (2018) A Tour of Unsupervised Deep Learning for Medical Image Analysis. ArXiv181207715 Cs Eess 16. Liu S, See KC, Ngiam KY, Celi LA, Sun X, Feng M (2020) Reinforcement Learning for Clinical Decision Support in Critical Care: Comprehensive Review. J Med Internet Res 22:e18477 17. Ordóñez FJ, Roggen D (2016) Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition. Sensors. https://doi.org/10.3390/s16010115 18. Rajeev R, Samath JA, Karthikeyan NK (2019) An Intelligent Recurrent Neural Network with Long Short-Term Memory (LSTM) BASED Batch Normalization for Medical Image Denoising. J Med Syst 43:234 19. Yu Y, Si X, Hu C, Zhang J (2019) A Review of Recurrent Neural Networks: LSTM Cells and Network Architectures. Neural Comput 31:1235–1270 20. Kass RE, Amari S, Arai K, et al (2018) Computational Neuroscience: Mathematical and Statistical Perspectives. Annu Rev Stat Its Appl 5:183–214 21. Bauer P, Thorpe A, Brunet G (2015) The quiet revolution of numerical weather prediction. Nature 525:47–55 22. Einevoll GT, Destexhe A, Diesmann M, Grün S, Jirsa V, de Kamps M, Migliore M, Ness TV, Plesser HE, Schürmann F (2019) The Scientific Case for Brain Simulations. Neuron 102:735–744 23. Fan X, Markram H (2019) A Brief History of Simulation Neuroscience. Front Neuroinformatics 13:32 24. Lytton WW, Arle J, Bobashev G, Ji S, Klassen TL, Marmarelis VZ, Schwaber J, Sherif MA, Sanger TD (2017) Multiscale modeling in the clinic: diseases of the brain and nervous system. Brain Inform 4:219–230 25. Cofre R, Herzog R, Mediano P, Piccinini J, Rosas F, Perl Y, Tagliazucchi E (2020) Whole-brain models to explore altered states of consciousness from the bottom up. 26. Parker D, Srivastava V (2013) Dynamic systems approaches and levels of analysis in the nervous system. Front Physiol. https://doi.org/10.3389/fphys.2013.00015 27. Catterall WA, Raman IM, Robinson HPC, Sejnowski TJ, Paulsen O (2012) The Hodgkin-Huxley Heritage: From Channels to Circuits. J Neurosci 32:14064–14073 28. Isakovic J, Dobbs-Dixon I, Chaudhury D, Mitrecić D (2018) Modeling of inhomogeneous electromagnetic fields in the nervous system: a novel paradigm in understanding cell interactions, disease etiology and therapy. Sci Rep. https://doi.org/10.1038/s41598-018-31054-9 29. Stefanou SS, Kastellakis G, Poirazi P (2016) Creating and Constraining Compartmental Models of Neurons Using Experimental Data. In: Korngreen A (ed) Adv. Patch-Clamp Anal. Neurosci. Springer, New York, NY, pp 325–343

30. Seo H, Jun S (2018) Relation between the electric field and activation of cortical neurons in transcranial electrical stimulation. Brain Stimulat. https://doi.org/10.1016/j.brs.2018.11.004 31. Teeter C, Iyer R, Menon V, et al (2018) Generalized leaky integrate-and-fire models classify multiple neuron types. Nat Commun 9:709 32. Wang T, Yin L, Zou X, Shu Y, Rasch MJ, Wu S (2016) A Phenomenological Synapse Model for Asynchronous Neurotransmitter Release. Front Comput Neurosci. https://doi.org/10.3389/fncom.2015.00153 33. Mäki-Marttunen T, Iannella N, Edwards AG, Einevoll GT, Blackwell KT (2020) A unified computational model for cortical post-synaptic plasticity. eLife 9:e55714 34. Graupner M, Brunel N (2018) Modeling Synaptic Plasticity in Hippocampus: A Calcium-Based Approach. In: Cutsuridis V, Graham BP, Cobb S, Vida I (eds) Hippocampal Microcircuits Comput. Model. Resour. Book. Springer International Publishing, Cham, pp 615–644 35. Falcon M, Jirsa V, Solodkin A (2016) A new neuroinformatics approach to personalized medicine in neurology: The Virtual Brain. Curr Opin Neurol 29:1 36. El Houssaini K, Bernard C, Jirsa VK (2020) The Epileptor Model: A Systematic Mathematical Analysis Linked to the Dynamics of Seizures, Refractory Status Epilepticus, and Depolarization Block. eNeuro. https://doi.org/10.1523/ENEURO.0485-18.2019 37. Ableidinger M, Buckwar E, Hinterleitner H (2017) A Stochastic Version of the Jansen and Rit Neural Mass Model: Analysis and Numerics. J Math Neurosci 7:8 38. Stefanovski L, Triebkorn P, Spiegler A, Diaz-Cortes M-A, Solodkin A, Jirsa V, McIntosh AR, Ritter P, Initiative for the ADN (2019) Linking Molecular Pathways and Large-Scale Computational Modeling to Assess Candidate Disease Mechanisms and Pharmacodynamics in Alzheimer’s Disease. Front Comput Neurosci. https://doi.org/10.3389/fncom.2019.00054 39. Kilpatrick ZP (2015) Wilson-Cowan Model. In: Jaeger D, Jung R (eds) Encycl. Comput. Neurosci. Springer, New York, NY, pp 3159–3163 40. Lanillos P, Oliva D, Philippsen A, Yamashita Y, Nagai Y, Cheng G (2020) A review on neural network models of schizophrenia and autism spectrum disorder. Neural Netw 122:338–363 41. Schroll H, Hamker FH (2013) Computational models of basal-ganglia pathway functions: focus on functional neuroanatomy. Front Syst Neurosci. https://doi.org/10.3389/fnsys.2013.00122 42. Strzelczyk A, Griebel C, Lux W, Rosenow F, Reese J-P (2017) The Burden of Severely Drug- Refractory Epilepsy: A Comparative Longitudinal Evaluation of Mortality, Morbidity, Resource Use, and Cost Using German Health Insurance Data. Front Neurol 8:712 43. Jablensky A (2016) Psychiatric classifications: validity and utility. World Psychiatry 15:26–31

44. Bennabi D, Charpeaud T, Yrondi A, et al (2019) Clinical guidelines for the management of treatment-resistant depression: French recommendations from experts, the French Association for Biological Psychiatry and Neuropsychopharmacology and the fondation FondaMental. BMC Psychiatry 19:262 45. Ma Y, Rosenheck R, Ye B, Fan N, He H (2020) Effectiveness of electroconvulsive therapy in patients with “less treatment-resistant” depression by the Maudsley Staging Model. Brain Behav 10:e01654 46. Tysnes O-B, Storstein A (2017) Epidemiology of Parkinson’s disease. J Neural Transm Vienna Austria 1996 124:901–905 47. Caligiore D, Helmich RC, Hallett M, Moustafa AA, Timmermann L, Toni I, Baldassarre G (2016) Parkinson’s disease as a system-level disorder. NPJ Park Dis 2:16025 48. Tambasco N, Romoli M, Calabresi P (2018) Levodopa in Parkinson’s Disease: Current Status and Future Developments. Curr Neuropharmacol 16:1239–1252 49. Baston C, Contin M, Calandra Buonaura G, Cortelli P, Ursino M (2016) A Mathematical Model of Levodopa Medication Effect on Basal Ganglia in Parkinson’s Disease: An Application to the Alternate Finger Tapping Task. Front Hum Neurosci. https://doi.org/10.3389/fnhum.2016.00280 50. Gao L, Zhang J, Chan P, Wu T (2016) Levodopa Effect on Basal Ganglia Motor Circuit in Parkinson’s Disease. CNS Neurosci Ther 23:76–86 51. Humphries MD, Obeso J, Dreyer JK (2018) Insights into Parkinson’s disease from computational models of the basal ganglia. J Neurol Neurosurg Psychiatry 89:1181–1188 52. Muddapu VR, Mandali A, Chakravarthy VS, Ramaswamy S (2019) A Computational Model of Loss of Dopaminergic Cells in Parkinson’s Disease Due to Glutamate-Induced Excitotoxicity. Front Neural Circuits. https://doi.org/10.3389/fncir.2019.00011 53. Frey A-L, McCabe C (2020) Effects of serotonin and dopamine depletion on neural prediction computations during social learning. Neuropsychopharmacology 45:1431–1437 54. Frank MJ, Samanta J, Moustafa AA, Sherman SJ (2007) Hold your horses: impulsivity, deep brain stimulation, and medication in parkinsonism. Science 318:1309–1312 55. Hartmann CJ, Fliegen S, Groiss SJ, Wojtecki L, Schnitzler A (2019) An update on best practice of deep brain stimulation in Parkinson’s disease. Ther Adv Neurol Disord 12:175628641983809 56. Bilge MT, Gosai AK, Widge AS (2018) Deep Brain Stimulation in Psychiatry. Psychiatr Clin North Am 41:373–383 57. Clair A-H, Haynes W, Mallet L (2018) Recent advances in deep brain stimulation in psychiatric disorders. F1000Research. https://doi.org/10.12688/f1000research.14187.1 58. Sturm V, Lenartz D, Koulousakis A, Treuer H, Herholz K, Klein JC, Klosterkötter J The Nucleus Accumbens: A Target for Deep- Brain Stimulation in Obsessive-Compulsive and Anxiety Disorders. 6

59. Zangiabadi N, Ladino LD, Sina F, Orozco-Hernández JP, Carter A, Téllez-Zenteno JF (2019) Deep Brain Stimulation and Drug-Resistant Epilepsy: A Review of the Literature. Front Neurol. https://doi.org/10.3389/fneur.2019.00601 60. Lozano AM, Lipsman N, Bergman H, et al (2019) Deep brain stimulation: current challenges and future directions. Nat Rev Neurol 15:148–160 61. Herrington TM, Cheng JJ, Eskandar EN (2016) Mechanisms of deep brain stimulation. J Neurophysiol 115:19–38 62. Jakobs M, Fomenko A, Lozano AM, Kiening KL (2019) Cellular, molecular, and clinical mechanisms of action of deep brain stimulation—a systematic review on established indications and outlook on future developments. EMBO Mol Med. https://doi.org/10.15252/emmm.201809575 63. Milosevic L, Kalia SK, Hodaie M, Lozano AM, Popovic MR, Hutchison WD (2018) Physiological mechanisms of thalamic ventral intermediate nucleus stimulation for tremor suppression. Brain 141:2142–2155 64. Creed M, Pascoli VJ, Lüscher C (2015) Addiction therapy. Refining deep brain stimulation to emulate optogenetic treatment of synaptic pathology. Science 347:659–664 65. Rughani A, Schwalb JM, Sidiropoulos C, et al (2018) Congress of Neurological Surgeons Systematic Review and Evidence-Based Guideline on Subthalamic Nucleus and Globus Pallidus Internus Deep Brain Stimulation for the Treatment of Patients With Parkinson’s Disease: Executive Summary. Neurosurgery 82:753–756 66. Mao Z, Ling Z, Pan L, Xu X, Cui Z, Liang S, Yu X (2019) Comparison of Efficacy of Deep Brain Stimulation of Different Targets in Parkinson’s Disease: A Network Meta-Analysis. Front Aging Neurosci. https://doi.org/10.3389/fnagi.2019.00023 67. Cury R, Fraix V, Castrioto A, Fernández M, Krack P, Chabardes S, Seigneuret E, Benabid A, Moro E, Alho E (2017) Thalamic deep brain stimulation for tremor in Parkinson disease, essential tremor, and dystonia. Neurology 89:10.1212/WNL.0000000000004295 68. Wang J-W, Zhang Y-Q, Zhang X-H, Wang Y-P, Li J-P, Li Y-J (2017) Deep Brain Stimulation of Pedunculopontine Nucleus for Postural Instability and Gait Disorder After Parkinson Disease: A Meta- Analysis of Individual Patient Data. World Neurosurg 102:72–78 69. Yu Y, Wang X, Wang Q, Wang Q (2020) A review of computational modeling and deep brain stimulation: applications to Parkinson’s disease. Appl Math Mech 41:1747–1768 70. Maling N, Lempka SF, Blumenfeld Z, Bronte-Stewart H, McIntyre CC (2018) Biophysical basis of subthalamic local field potentials recorded from deep brain stimulation electrodes. J Neurophysiol 120:1932–1944

71. Yousif N, Bain PG, Nandi D, Borisyuk R (2020) A Population Model of Deep Brain Stimulation in Movement Disorders From Circuits to Cells. Front Hum Neurosci 14:55 72. Meijer H, Krupa M, Cagnan H, Lourens M, Heida T, Martens H, Gils S (2011) From Parkinsonian thalamic activity to restoring thalamic relay using deep brain stimulation: New insights from computational modeling. J Neural Eng 8:066005 73. Karamintziou SD, Deligiannis NG, Piallat B, et al (2015) Dominant efficiency of nonregular patterns of subthalamic nucleus deep brain stimulation for Parkinson’s disease and obsessive-compulsive disorder in a data-driven computational model. J Neural Eng 13:016013 74. Paff M, Loh A, Sarica C, Lozano AM, Fasano A (2020) Update on Current Technologies for Deep Brain Stimulation in Parkinson’s Disease. J Mov Disord 13:185–198 75. Ughratdar I, Samuel M, Ashkan K (2015) Technological Advances in Deep Brain Stimulation. J Park Dis 5:483–496 76. Miocinovic S, Ostrem JL, Okun MS, Bullinger KL, Riva-Posse P, Gross RE, Buetefisch CM (2020) Recommendations for Deep Brain Stimulation Device Management During a Pandemic. J Park Dis 10:903–910 77. Reich MM, Steigerwald F, Sawalhe AD, Reese R, Gunalan K, Johannes S, Nickl R, Matthies C, McIntyre CC, Volkmann J (2015) Short pulse width widens the therapeutic window of subthalamic neurostimulation. Ann Clin Transl Neurol 2:427–432 78. Parker T, Raghu ALB, FitzGerald JJ, Green AL, Aziz TZ (2020) Multitarget deep brain stimulation for clinically complex movement disorders. J Neurosurg 134:351–356 79. Thenaisie Y, Palmisano C, Canessa A, et al (2021) Towards adaptive deep brain stimulation: clinical and technical notes on a novel commercial device for chronic brain sensing. medRxiv 2021.03.10.21251638 80. Adamchic I, Hauptmann C, Barnikol UB, et al (2014) Coordinated reset neuromodulation for Parkinson’s disease: Proof-of-concept study. Mov Disord 29:1679–1684 81. Fleming JE, Dunn E, Lowery MM (2020) Simulation of Closed-Loop Deep Brain Stimulation Control Schemes for Suppression of Pathological Beta Oscillations in Parkinson’s Disease. Front Neurosci. https://doi.org/10.3389/fnins.2020.00166 82. Kokkinos V, Sisterson ND, Wozny TA, Richardson RM (2019) Association of Closed-Loop Brain Stimulation Neurophysiological Features With Seizure Control Among Patients With Focal Epilepsy. JAMA Neurol 76:800 83. Meier JM, Perdikis D, Blickensdörfer A, Stefanovski L, Liu Q, Maith O, Dinkelbach HÜ, Baladron J, Hamker FH, Ritter P (2021) Virtual deep brain stimulation: Multiscale co-simulation of a spiking

50 basal ganglia model and a whole-brain mean-field model with The Virtual Brain. bioRxiv 2021.05.05.442704 84. Petersen MV, Mlakar J, Haber SN, Parent M, Smith Y, Strick PL, Griswold MA, McIntyre CC (2019) Holographic Reconstruction of Axonal Pathways in the Human Brain. Neuron 104:1056-1064.e3 85. Fisher RS, Cross JH, D’Souza C, et al (2017) Instruction manual for the ILAE 2017 operational classification of seizure types. Epilepsia 58:531–542 86. Beghi E (2020) The Epidemiology of Epilepsy. Neuroepidemiology 54:185–191 87. Singh SP, Sankaraneni R, Antony AR (2017) Evidence-based guidelines for the management of epilepsy. Neurol India 65:S6–S11 88. Dalic L, Cook M (2016) Managing drug-resistant epilepsy: challenges and solutions. Neuropsychiatr Dis Treat Volume 12:2605–2616 89. Golyala A, Kwan P (2017) Drug development for refractory epilepsy: The past 25 years and beyond. Seizure 44:147–156 90. Mohan M, Keller S, Nicolson A, Biswas S, Smith D, Osman Farah J, Eldridge P, Wieshmann U (2018)The long-term outcomes of epilepsy surgery. PLoS ONE. https://doi.org/10.1371/journal.pone.0196274 91. Sheng J, Liu S, Qin H, Li B, Zhang X (2018) Drug-Resistant Epilepsy and Surgery. Curr Neuropharmacol 16:17–28 92. Holdgraf C, Appelhoff S, Bickel S, et al (2019) iEEG-BIDS, extending the Brain Imaging Data Structure specification to human intracranial electrophysiology. Sci Data. https://doi.org/10.1038/s41597-019-0105-7 93. Mikulan E, Russo S, Parmigiani S, et al (2020) Simultaneous human intracerebral stimulation and HD-EEG, ground-truth for source localization methods. Sci Data. https://doi.org/10.1038/s41597-020- 0467-x 94. Almustafa KM (2020) Classification of epileptic seizure dataset using different machine learning algorithms. Inform Med Unlocked 21:100444 95. Raghu null, Sriraam N, Temel Y, Rao SV, Kubben PL (2019) A convolutional neural network based framework for classification of seizure types. Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Int Conf 2019:2547–2550 96. Maksimenko VA, Kurkin SA, Pitsik EN, Musatov VY, Runnova AE, Efremova TY, Hramov AE, Pisarchik AN (2018) Artificial Neural Network Classification of Motor-Related EEG: An Increase in Classification Accuracy by Reducing Signal Complexity. Complexity 2018:e9385947 97. Abdelhameed A, Bayoumi M (2021) A Deep Learning Approach for Automatic Seizure Detection in Children With Epilepsy. Front Comput Neurosci. https://doi.org/10.3389/fncom.2021.650050

98. Elahian B, Yeasin M, Mudigoudar B, Wheless JW, Babajani-Feremi A (2017) Identifying seizure onset zone from electrocorticographic recordings: A machine learning approach based on phase locking value. Seizure 51:35–42 99. Boon P, De Cock E, Mertens A, Trinka E (2018) Neurostimulation for drug-resistant epilepsy: a systematic review of clinical evidence for efficacy, safety, contraindications and predictors for response. Curr Opin Neurol 31:198–210 100. Fisher B, DesMarteau JA, Koontz EH, Wilks SJ, Melamed SE (2021) Responsive Vagus Nerve Stimulation for Drug Resistant Epilepsy: A Review of New Features and Practical Guidance for Advanced Practice Providers. Front Neurol. https://doi.org/10.3389/fneur.2020.610379 101. Hartshorn A, Jobst B (2018) Responsive brain stimulation in epilepsy. Ther Adv Chronic Dis 9:135–142 102. González HFJ, Yengo-Kahn A, Englot DJ (2019) Vagus Nerve Stimulation for the Treatment of Epilepsy. Neurosurg Clin N Am 30:219–230 103. Sandler RA, Song D, Hampson RE, Deadwyler SA, Berger TW, Marmarelis VZ (2015) Hippocampal closed-loop modeling and implications for seizure stimulation design. J Neural Eng 12:056017 104. Raikov I, Soltesz I (2015) A Master Plan for the Epilepsies? Toward a General Theory of Seizure Dynamics. Epilepsy Curr 15:133–135 105. Jirsa VK, Stacey WC, Quilichini PP, Ivanov AI, Bernard C (2014) On the nature of seizure dynamics. Brain 137:2210–2230 106. Saggio ML, Crisp D, Scott JM, et al A taxonomy of seizure dynamotypes. eLife. https://doi.org/10.7554/eLife.55632 107. Jirsa VK, Proix T, Perdikis D, et al (2017) The Virtual Epileptic Patient: Individualized whole- brain models of epilepsy spread. NeuroImage 145:377–388 108. Nichols E, Szoeke CEI, Vollset SE, et al (2019) Global, regional, and national burden of Alzheimer’s disease and other dementias, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol 18:88–106 109. Yiannopoulou KG, Papageorgiou SG (2020) Current and Future Treatments in Alzheimer Disease: An Update. J Cent Nerv Syst Dis. https://doi.org/10.1177/1179573520907397 110. Pellegrini E, Ballerini L, Hernandez MDCV, et al (2018) Machine learning of neuroimaging for assisted diagnosis of cognitive impairment and dementia: A systematic review. Alzheimers Dement Amst Neth 10:519–535 111. Domingo-Fernández D, Kodamullil AT, Iyappan A, Naz M, Emon MA, Raschka T, Karki R, Springstubbe S, Ebeling C, Hofmann-Apitius M (2017) Multimodal mechanistic signatures for

52 neurodegenerative diseases (NeuroMMSig): a web server for mechanism enrichment. Bioinformatics 33:3679–3681 112. Liu P-P, Xie Y, Meng X-Y, Kang J-S (2019) History and progress of hypotheses and clinical trials for Alzheimer’s disease. Signal Transduct Target Ther 4:1–22 113. Du X, Wang X, Geng M (2018) Alzheimer’s disease hypothesis and related therapies. Transl Neurodegener. https://doi.org/10.1186/s40035-018-0107-y 114. Makin S (2018) The amyloid hypothesis on trial. Nature 559:S4–S7 115. Geerts H, Hofmann-Apitius M, Anastasio TJ, Brain Health Modeling Initiative (2017) Knowledge- driven computational modeling in Alzheimer’s disease research: Current state and future trends. Alzheimers Dement J Alzheimers Assoc 13:1292–1302 116. Hughes RE, Nikolic K, Ramsay RR (2016) One for All? Hitting Multiple Alzheimer’s Disease Targets with One Drug. Front Neurosci. https://doi.org/10.3389/fnins.2016.00177 117. Stefanovski L, Meier JM, Pai RK, et al (2021) Bridging Scales in Alzheimer’s Disease: Biological Framework for Brain Simulation With The Virtual Brain. Front Neuroinformatics 15:630172 118. Reas ET (2017) Amyloid and Tau Pathology in Normal Cognitive Aging. J Neurosci 37:7561–7563 119. Wafa Hatem A., Wolfe Charles D.A., Emmett Eva, Roth Gregory A., Johnson Catherine O., Wang Yanzhong (2020) Burden of Stroke in Europe. Stroke 51:2418–2427 120. Reinkensmeyer DJ, Burdet E, Casadio M, Krakauer JW, Kwakkel G, Lang CE, Swinnen SP, Ward NS, Schweighofer N (2016) Computational neurorehabilitation: modeling plasticity and learning to predict recovery. J NeuroEngineering Rehabil 13:42 121. Cramer SC, Wolf SL, Adams HP, et al (2017) Stroke Recovery and Rehabilitation Research: Issues, Opportunities, and the National Institutes of Health StrokeNet. Stroke 48:813–819 122. Falcon MI, Riley JD, Jirsa V, McIntosh AR, Shereen AD, Chen EE, Solodkin A (2015) The Virtual Brain: Modeling Biological Correlates of Recovery after Chronic Stroke. Front Neurol. https://doi.org/10.3389/fneur.2015.00228 123. Falcon MI, Riley JD, Jirsa V, McIntosh AR, Chen EE, Solodkin A (2016) Functional Mechanisms of Recovery after Chronic Stroke: Modeling with the Virtual Brain. eNeuro. https://doi.org/10.1523/ENEURO.0158-15.2016 124. Cuthbert BN (2015) Research Domain Criteria: toward future psychiatric nosologies. Dialogues Clin Neurosci 17:89–97 125. Montague PR, Dolan RJ, Friston KJ, Dayan P (2012) Computational psychiatry. Trends Cogn Sci 16:72–80 126. Adams RA, Huys QJM, Roiser JP (2015) Computational Psychiatry: towards a mathematically informed understanding of mental illness. J Neurol Neurosurg Psychiatry jnnp-2015-310737

127. Yassin W, Nakatani H, Zhu Y, et al (2020) Machine-learning classification using neuroimaging data in schizophrenia, autism, ultra-high risk and first-episode psychosis. Transl Psychiatry 10:1–11 128. Khodayari-Rostamabad A, Reilly JP, Hasey GM, de Bruin H, MacCrimmon DJ (2013) A machine learning approach using EEG data to predict response to SSRI treatment for major depressive disorder. Clin Neurophysiol 124:1975–1985 129. Balu DT (2016) The NMDA Receptor and Schizophrenia: From Pathophysiology to Treatment. Adv Pharmacol San Diego Calif 76:351–382 130. Mäki-Marttunen T, Kaufmann T, Elvsåshagen T, et al (2019) Biophysical Psychiatry—How Computational Neuroscience Can Help to Understand the Complex Mechanisms of Mental Disorders. Front Psychiatry. https://doi.org/10.3389/fpsyt.2019.00534 131. Jackevičius R, Graham BP, Saudargienė A (2017) Influence of NMDA and GABA synaptic dysfunction on the evoked gamma oscillations in a computational model of schizophrenia. Biologija. https://doi.org/10.6001/biologija.v63i2.3532 132. Naples JG, Ruckenstein MJ (2020) Cochlear Implant. Otolaryngol Clin North Am 53:87–102 133. Benabid AL, Costecalde T, Eliseyev A, et al (2019) An exoskeleton controlled by an epidural wireless brain–machine interface in a tetraplegic patient: a proof-of-concept demonstration. Lancet Neurol 18:1112–1122 134. Musk E, Neuralink (2019) An Integrated Brain-Machine Interface Platform With Thousands of Channels. J Med Internet Res 21:e16194 135. Edelman B, Johnson N, Sohrabpour A, Tong S, Thakor N, He B (2015) Systems Neuroengineering: Understanding and Interacting with the Brain. Engineering 1:292 136. Baston C, Ursino M (2015) A Biologically Inspired Computational Model of Basal Ganglia in Action Selection. Comput Intell Neurosci. https://doi.org/10.1155/2015/187417 137. Wilson HR, Cowan JD (1972) Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons. Biophys J 12:1–24 138. Bressloﬀ PC Metastable states and quasicycles in a stochastic Wilson-Cowan model of neuronal population dynamics. 22 139. Beudel M, Sadnicka A, Edwards M, de Jong BM (2019) Linking Pathological Oscillations With Altered Temporal Processing in Parkinsons Disease: Neurophysiological Mechanisms and Implications for Neuromodulation. Front Neurol. https://doi.org/10.3389/fneur.2019.00462 140. Halje P, Brys I, Mariman JJ, da Cunha C, Fuentes R, Petersson P (2019) Oscillations in cortico- basal ganglia circuits: implications for Parkinson’s disease and other neurologic and psychiatric conditions. J Neurophysiol 122:203–231

141. McDermott B, Porter E, Hughes D, McGinley B, Lang M, O’Halloran M, Jones M Gamma Band Neural Stimulation in Humans and the Promise of a New Modality to Prevent and Treat Alzheimer’s Disease. J Alzheimers Dis 65:363–392 142. Koelman LA, Lowery MM (2019) Beta-Band Resonance and Intrinsic Oscillations in a Biophysically Detailed Model of the Subthalamic Nucleus-Globus Pallidus Network. Front Comput Neurosci. https://doi.org/10.3389/fncom.2019.00077 143. Mirdamadi JL (2016) Cerebellar role in Parkinson’s disease. J Neurophysiol 116:917–919 144. Mueller K, Jech R, Ballarini T, Holiga Š, Růžička F, Piecha FA, Möller HE, Vymazal J, Růžička E, Schroeter ML (2019) Modulatory Effects of Levodopa on Cerebellar Connectivity in Parkinson’s Disease. Cerebellum Lond Engl 18:212–224 145. Dong J, Hawes S, Wu J, Le W, Cai H (2021) Connectivity and Functionality of the Globus Pallidus Externa Under Normal Conditions and Parkinson’s Disease. Front Neural Circuits. https://doi.org/10.3389/fncir.2021.645287 146. Dayal V, Limousin P, Foltynie T (2017) Subthalamic Nucleus Deep Brain Stimulation in Parkinson’s Disease: The Effect of Varying Stimulation Parameters. J Park Dis 7:235–245 147. Zibetti M, Moro E, Krishna V, Sammartino F, Picillo M, Munhoz RP, Lozano AM, Fasano A (2016) Low-frequency Subthalamic Stimulation in Parkinson’s Disease: Long-term Outcome and Predictors. Brain Stimulat 9:774–779 148. Bouthour W, Wegrzyk J, Momjian S, et al (2018) Short pulse width in subthalamic stimulation in Parkinson’s disease: a randomized, double-blind study: Short Pulse Width in DBS. Mov Disord 33:169– 173 149. Neumann W-J, Staub F, Horn A, Schanda J, Mueller J, Schneider G-H, Brown P, Kühn AA (2016) Deep brain recordings using an implanted pulse generator in Parkinson’s disease. Neuromodulation J Int Neuromodulation Soc 19:20–24

15. ANNEXES

1. PROJECT PARTICIPATION

Prediction of neurosurgical treatment outcomes in Parkinson’s disease (The Human Brain Project EBRAINS Research Infrastructure Voucher Programme Call 2020), 2021-2022.

PI: Prof. Aušra Saudargienė; Human Brain Project collaborator: Prof. Viktor Jirsa.

Institut de Neurosciences des Systèmes, Aix-Marseille Université Faculté de Médecine, Marseille, France.

2. ATTENDED SEMINARS

#1 Brain Matters | 21 September 2020 | 16:00–17:00 CEST

EBRAINS – in Search of Breakthroughs in Science and Medicine

Speakers: Viktor Jirsa (Aix-Marseille University), Timo Dickscheid (Forschungszentrum Jülich) , Marmaduke Woodman (Aix-Marseille University), Sandra Diaz Pier (Forschungszentrum Jülich)

Sessions: The Human Brain Project and EBRAINS: A new era in brain research, Using Brain Scan Data to Predict Personality Traits and Pathologies, Guiding Medical Interventions with Personalized Brain Models https://www.humanbrainproject.eu/en/brain-matters/

#2 Brain Matters | 27 October 2020 | 16:00–17:00 CET

Networks for Consciousness & Medical Use Cases

Moderator: Katrina Sichel

Speakers: Mavi Sanchez-Vives (Consorci Institut d'Investigacions Biomediques August Pi i Sunyer), Marcello Massimini (University of Milan), Pieter Roelfsema (Netherlands Institute for Neuroscience)

Sessions: “Brain Networks Underlying Cognition and Consciousness”, “Determining consciousness in non-responsive or minimally-responsive patients”. How can we calibrate computer brain interfaces to ensure electrical stimuli inputs reach the level of attention?” https://www.humanbrainproject.eu/en/brain-matters/

#3 Brain Matters | 10 December 2020 | 16:00–17:00 CET

Brain Inspired Technology and Architectures

Speakers: Steve Furber (University of Manchester), Rainer Goebel (Maastricht University)

Christopher Summerfield (University of Oxford)

Sessions: The origins of bio-inspired computing, "Adaptive networks for cognitive architectures: from advanced learning to neurorobotics and neuromorphic applications", Efforts to build better artificial intelligence. https://www.humanbrainproject.eu/en/brain-matters/

#4 Brain Matters | 25 February 2021 | 16:00–17:30 CET

Brain Simulation Science and Technology: looking forward

Speakers: Jeanette Hellgren Kotaleski (KTH Royal Institute of Technology), Eduardo Ros (University of Granada) Michele Migliore (National Research Council of Italy).

Sessions: Models of Hippocampus, Cerebellum and Basal ganglia. https://www.humanbrainproject.eu/en/brain-matters/

#5 Brain Matters | 7 May 2021 | 12:00–13:00 CEST

Introduction to The EBRAINS Virtual Big Brain

Speakers: Viktor Jirsa (Aix-Marseille University), Timo Dickscheid (Forschungszentrum Jülich), Marmaduke Woodman (Aix-Marseille University), Sandra Diaz Pier (Forschungszentrum Jülich)

Description: In the fifth edition of the Brain Matters Webinar we turned our attention from scientific results to scientific activity underway in the final phase of the Human Brain Project. In the Virtual Big Brain project teams at Forschungszentrum Jülich in Germany and Aix-Marseille University are working together to build a massively scaled up version of The Virtual Brain (TVB) a tool which simulates the whole human brain using mean field models. https://www.humanbrainproject.eu/en/brain-matters/

3. COMPLETED COURSES

I. Modelling Strokes within TVB II. The Local Epileptor: Parts 1 & 2 III. The Virtual Epileptic Patient: Parts 1 & 2 IV. TVB-NEST co-simulation on local computer, Human Brain Project (HBP) TVB-NEST co- simulation V. Human Brain Project (HBP) image processing pipeline for the Virtual Brain VI. An automated pipeline for constructing personalized virtual brains from multimodal neuroimaging data VII. Modeling brain dynamics in brain tumor patients using The Virtual Brain VIII. Linking molecular pathways and large-scale computational modeling to assess candidate disease mechanisms and pharmacodynamics in Alzheimer’s disease IX. Inferring multi-scale neural mechanisms with brain network modelling

The Virtual Brain Simulation Platform:

https://training.incf.org/