Xai-EWS — an Explainable AI Model Predicting Acute Critical Illness

Health & Medicine ︱Simon Meyer Lauritsen Patient data Prediction module Output P-Creatinine Acute Kidney

Injury: l

Kidney eGFR e d

o 78% m

P-Potassium

t .. .

xAI-EWS — an explainable u

n i AI model predicting P-Albumin time Patient data with explanations Explanation module Why? https://www.nature.com/articles/s41467-020-17431-x P-Creatinine n acute critical illness PhotoCredit: Kidney eGFR o at i

P-Potassium n

he time it takes medical staff to choices of patients, providers, payers, .. . Early clinical predictions of xpl a acute critical illness have a vital diagnose a patient’s acute critical etc. to achieve better outcomes in e influence on patient outcomes. Tillness has a pivotal influence terms of quality of care and its cost. P-Albumin For clinical medicine to benefit on the patient outcome. Acute critical With such high-stake applications, the

from the higher predictive illness usually follows a deterioration simpler, more transparent systems are Low parameter value High Low parameter relevance High power of Artificial Intelligence, of the patient’s vital signs – the often chosen so that clinicians can easily however, explainable and routinely measured heart rate, body back-trace a prediction. This strategy, Overview of the xAI-EWS system. Based on patient data, the model makes a prediction, such as a 78% risk of acute kidney injury. The explanation transparent systems are temperature, respiration rate and blood however, can lead to negative outcomes module then explains the predictions in terms of input variables, e.g. p-creatinine, kidney eGFR, p-potassium and p-albumin. essential. Simon Meyer pressure. Early clinical predictions tend for patients. Lauritsen and his collaborators to be based on manual calculations performance, providing early, real- at Enversion, Aarhus University, using these clinical parameters that Simon Meyer Lauritsen, a Biomedical time prediction of acute critical illness. For clinical medicine to benefit from and Regional Hospital Horsens produce screening metrics, including Engineer and Industrial PhD researcher Unfortunately, the lack of information the higher predictive power of AI, in Denmark have developed early warning score (EWS) systems at Enversion A/S and Aarhus University, regarding the complex decisions xAI-EWS – an explainable AI such as the Sequential Organ Failure together with his co-workers, has made by such systems hindered their explainable and transparent Deep early warning score system for Assessment score (SOFAs). developed a robust and accurate AI clinical translation. the prediction of acute critical model capable of predicting if a patient Learning algorithms are crucial. illness using electronic While Artificial Intelligence (AI) lends will develop an acute critical illness. EXPLAINABLE AI EARLY WARNING health records. itself to producing earlier predictions Moreover, they have designed this SCORE SYSTEM: XAI-EWS biochemistry, medicine, microbiology, surgery were among the various with greater accuracy than the algorithm in such a way that it supports Using historical health data in the and procedure codes. It was extracted departments in each of the hospitals. traditional EWS systems, these black- the clinician by providing an explanation form of electronic health records, the from the CROSS-TRACKS cohort, a box predictions are not easily explained of its prediction with reference to research team’s innovative AI algorithm population-based Danish cross-sectorial Dr Lauritsen extracted data on all to clinicians. This means a trade-off the electronic health record data can predict whether a patient will cohort comprising a mixed rural and 163,050 available inpatient admissions between transparency and the potential supporting it. develop an acute critical illness. This urban multicentre population served by from this period for inclusion in his power of predictive medicine takes algorithm has been trained to recognise four regional hospitals as well as one analysis. A total of 66,288 individual place. Healthcare transparency involves DEEP LEARNING ALGORITHMS cases from the historical data that are larger university hospital. Emergency patients were involved in these making information on quality, efficiency Dr Lauritsen and the team explain that similar to the current case. Given the medicine, intensive care, and thoracic admissions. The patients had an and patients experience available to in order for clinical medicine to benefit patient’s vital parameters and blood average age of 55 years and 45.9% of the public. This must be reliable and from the higher predictive power of tests, it can identify the early signs of them were male. The study focused on The AI algorithm can predict whether a comprehensive in order to inform the AI, explainable and transparent Deep critical illness. It then presents the user patient will develop an acute critical illness three acute critical illnesses frequently Learning algorithms are crucial. Deep with an assessment of when and how to based on historical health data. observed in emergency medicine cases: Learning is a machine learning method proceed with the patient’s care. sepsis, acute kidney injury, and acute that is particularly suitable for big lung injury. The prevalence of these data sets. Deep Learning algorithms The researchers demonstrate how the emergency medicine cases among deploy artificial neural networks based explainable AI early warning score the admissions was 2.44%, 0.75%, on the structure and function of the system, xAI-EWS in short, overcomes and 1.68%, respectively. The model brain and take advantage of copious the shortfalls observed in other parameters comprised 27 laboratory metamorworks/Shutterstock.com cheap computation. These algorithms computational models by taking a parameters and six vital signs. These involve large complex neural networks, Deep Learning approach to analysing a features were selected by clinicians and their performance will continue to diverse multicentre data set. They used specialising in emergency medicine. increase as they are trained with more the secondary healthcare data of all and more data. residents aged 18 years and over in four PREDICTIVE PERFORMANCE Danish municipalities collected over the The predictive performance of this Previous research into electronic five-year period from 2012 to 2017. This digital medicine model was measured health records-trained AI systems has data was made up of information from in comparison with three baseline demonstrated high levels of predictive electronic health records and included prediction models for acute critical illness: angellodeco/Shutterstock.com

www.researchoutreach.org www.researchoutreach.org Behind the Research Halfpoint/Shutterstock.com

Dr Simon Meyer Dr Bo Dr Jeppe Dr Marianne Lauritsen Thiesson Lange J. Jørgensen

E: [email protected] T: +45 30554920 https://www.linkedin.com/in/simonlauritsen/ Research Objectives

Simon Meyer Lauritsen and his co-workers developed an AI algorithm that can predict if a patient will develop an acute critical illness – and explain why. In relation to the Covid-19 pandemic, xAI-EWS can help doctors assess if or when a patient will need external ventilation. Detail

the traditional models MEWS (modified understand the underlying reasoning In relation to the current Covid-19 Simon Meyer Lauritsen topic of applied machine learning. Marianne J. Jørgensen is Research early warning score), SOFA, and another behind the predictions. pandemic, xAI-EWS can help doctors Fiskerivej 12 1 floor Manager, Department of Research, AI model, a machine-learning algorithm assess if or when a patient will need 8000 Aarhus C Bo Thiesson is Chief Technology Regional Hospital Horsens. that has had great results in a previous The researchers emphasise another external ventilation. While Covid-19 is Denmark Officer and Analytical Director at randomised study. significant benefit of the xAI-EWS a new disease, impaired lung function Enversion A/S and kaunt A/S. He is Funding system: the daily workflow of healthcare is well known, so predictions based on Bio also Honorary Professor at Aarhus Innovation Fund Denmark The clinical usefulness of the xAI-EWS professionals at hospitals does not previous data on impaired lung function Simon Meyer Lauritsen is a University. system was evaluated by calculating have to be altered to take advantage can contribute to the fast and effective Biomedical Engineer, Paramedic, and Collaborators the net benefit at varying decision of the system’s predictive power, as the treatment for the patient. Industrial PhD researcher at Enversion Jeppe Lange is Director of Medical • Enversion thresholds for the A/S and Aarhus University in Denmark. Research, Associate Professor, • The Horsens Regional Hospital model, based on The concept of He has authored a book chapter on Consultant Orthopaedic Surgeon • MedTech Innovation Consortium the number of true AI can attract artificial intelligence in medicine as at Regional Hospital Horsens and (MTIC) The greatest value of detecting critical well as several publications on the Aarhus University. • Aarhus University positives and false scepticism. How positives generated illness is that potential interventions can we ensure from a sample of that an algorithm References Personal Response cases. This novel become timelier and more focused. is actually doing approach for early what we want Lauritsen, S.M., Kristensen, M., Olsen, M.V. et al. (2020). detection had more true positives and model is based on data that is routinely it to? In this case, the critical illness Explainable artificial intelligence model to predict acute What inspired you to develop the xAI-EWS system? fewer false negatives than any of the captured in the electronic health prediction algorithm ensures that the critical illness from electronic health records. Nature I have a background in prehospital emergency care, existing alarm systems. Throughout this records. Furthermore, while clinicians physician is involved throughout the Communications, [online] 11, 3852. Available at: https:// and I’m married to a medical doctor who has worked in comparative study the xAI-EWS system are entering the raw patient data into process. A crucial part of the research doi.org/10.1038/s41467-020-17431-x [Accessed 24th the emergency room. We have both experienced how critical conditions can develop quickly and surprise even outperformed all three baseline models. the electronic health records, xAI-EWS team’s work on the development of the August 2020] experienced clinicians. Sepsis alone is the third most makes real-time predictions available, xAI-EWS system is xAI – “Explainable common cause of death globally and contributes to 15% QUALITY OF THE EXPLANATIONS improving the timeliness of detection of artificial intelligence”. From a clinical Lauritsen, S.M., Kalør, M.E., Kongsgaard, E.L. et al. of deaths in Denmark. When I was introduced to the The quality of the explanations acute critical illnesses. implementation perspective, it is essential (2020). Early detection of sepsis utilizing deep learning possibilities of applying machine learning to health data, accompanying the xAI-EWS predictions that the algorithm can accurately explain on electronic health record event sequences. Artificial it was clear to me that these techniques could make a was also assessed by manual BROADER IMPACT its recommendations and supports Intelligence in Medicine, [online] 104, 101820, Available at: difference for sick people in need. inspection carried out by clinicians The researchers explain that “the the clinical staff with explanations of its https://doi.org/10.1016/j.artmed.2020.101820 [Accessed specialising in emergency medicine. greatest value of detecting critical prediction. The research team concludes, 24th August 2020] The appraisal of the predictions and illness is that potential interventions “in this way, doctors can act on the explanations generated by the xAI- become timelier and more focused”. detection and ensure that patients EWS system demonstrate its general This means that resources for both receive the right treatment. But at the clinical relevance. In addition to its treatment and care can be better same time [they] also fully understand high predictive performance, xAI-EWS utilised. Moreover, the affected patients why the algorithm has calculated a risk. provides explanations for the predictions benefit from an improved chance This fundamental, and unique, feature in terms of pinpointing the decisive of survival, fewer side effects and of the algorithm is considered crucial to input data which enables clinicians to increased quality of life. ensure an optimal clinical translation.”