804 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 45, NO. 4, APRIL 2010 A Micro-Power EEG Acquisition SoC With Integrated Feature Extraction Processor for a Chronic Seizure Detection System Naveen Verma, Member, IEEE, Ali Shoeb, Jose Bohorquez, Member, IEEE, Joel Dawson, Member, IEEE, John Guttag, and Anantha P. Chandrakasan, Fellow, IEEE Abstract—This paper presents a low-power SoC that performs Brain monitoring thus introduces key challenges for electronic EEG acquisition and feature extraction required for contin- systems in terms of both instrumentation and information uous detection of seizure onset in epilepsy patients. The SoC extraction. corresponds to one EEG channel, and, depending on the pa- tient, up to 18 channels may be worn to detect seizures as part Seizure detection in epilepsy patients is an important appli- of a chronic treatment system. The SoC integrates an instru- cation that is representative of the challenges. This paper de- mentation amplifier, ADC, and digital processor that streams scribes the details of an SoC that performs feature extraction features-vectors to a central device where seizure detection is from an analog EEG channel into the digital domain; the output performed via a machine-learning classifier. The instrumen- is used to detect the onset of seizures by way of a machine- tation-amplifier uses chopper-stabilization in a topology that achieves high input-impedance and rejects large electrode-off- learning classifier that is trained to patient-specific data. EEG sets while operating at 1 V; the ADC employs power-gating for sensing is targeted so that the system is noninvasive. This im- low energy-per-conversion while using static-biasing for com- plies that microvolt signals must be acquired from electrodes parator precision; the EEG feature extraction processor employs having very poor output impedance while in the presence of low-power hardware whose parameters are determined through numerous physiological and environmental interferences (e.g., validation via patient data. The integration of sensing and local processing lowers system power by 14x by reducing the rate of EMG, hum, etc.). Further, for reliable detection, subtle patient- wireless EEG data transmission. Feature vectors are derived at a specific EEG signal correlations must be determined over mul- rate of 0.5 Hz, and the complete one-channel SoC operates from a tiple channels (up to 18). The following sections start by de- 1 V supply, consuming 9 J per feature vector. scribing the opportunity and algorithm approach for patient-spe- Index Terms—1/f noise, algorithm design and analysis, ampli- cific seizure detection. Then, the SoC is described from both fiers, biomedical equipment, brain, choppers, digital signal pro- the system perspective and the IC implementation perspective. cessing, electroencephalography, low-noise amplifiers, low-power Finally, IC results are described, followed by a system demon- electronics. stration and conclusions. I. INTRODUCTION II. EPILEPSY AND SEIZURE DETECTION ECENTLY therapeutic and prosthetic devices have Epilepsy is a neurological disorder that causes a recurring R begun emerging that hold great promise for the treatment abnormal firing in groups of neurons. As a result patients ex- of patients with neurological conditions ranging from epilepsy perience seizures causing loss of coherence/cognition, loss of [1], Parkinson’s disease [2], narcolepsy [3], depression [4], and motor control, involuntary motion (convulsions), and possibly motor impairments [5]. The ability to acquire targeted neuro- even death. Fig. 1 shows PET scan images highlighting a partic- logical information from the brain is an essential requirement ular firing pattern that is associated with seizures (ictal period) to the advancement of these systems. This implies the need to in the considered patient. The EEG during a seizure onset is also sense neural signals but, more critically, to use these in order to shown. Although EEG has the benefit that it is noninvasive, its establish correlation with the actual clinical states of interest. correlation with seizures is complicated by the attenuation, 1/f filtering, and spatial aliasing of the neural field potentials across Manuscript received August 24, 2009; revised January 15, 2010. Current ver- the skull and skin. Nonetheless, taking the recording in Fig. 1 sion published March 24, 2010. This paper was approved by Guest Editor Ajith as an example, approximately 7.5 sec before the start of clinical Amerasekera. The work of N. Verma was supported in part by the Intel Foun- symptoms, a subtle but characteristic change in the EEG can be dation Ph.D. Fellowship Program and NSERC. IC fabrication was provided by National Semiconductor Corporation. observed. If this electrical onset can be detected, an advanced N. Verma is with Princeton University, Princeton, NJ 08544 USA (e-mail: signal can be generated to warn the patient and caregivers, ac- [email protected]). tuate a therapeutic stimulator (e.g., [6], [7]), or trigger EEG data A. Shoeb, J. Bohorquez, J. Dawson, J. Guttag, and A. P. Chandrakasan are with the Massachusetts Institute of Technology, Cambridge, MA 02139 storage for analysis by a neurologist. USA (e-mail: [email protected]; [email protected]; [email protected]; Although the critical variances in the electrical onset are [email protected]; [email protected]). minute and variable from patient to patient, [8] shows that Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. seizures are stereotypical for a given patient. Machine learning Digital Object Identifier 10.1109/JSSC.2010.2042245 can thus be used to train a classifier on a patient-by-patient 0018-9200/$26.00 © 2010 IEEE Authorized licensed use limited to: MIT Libraries. Downloaded on April 10,2010 at 20:34:14 UTC from IEEE Xplore. Restrictions apply. VERMA et al.: EEG ACQUISITION SoC WITH INTEGRATED FEATURE EXTRACTION PROCESSOR FOR A CHRONIC SEIZURE DETECTION SYSTEM 805 The algorithm was validated through tests on 536 hours of data over 16 patients. Ref. [8] shows that the approach of patient-specific learning simultaneously improves sensitivity, specificity, and latency (the values achieved are 93%, 0.3 0.7 false alarms/hour, and 6.7 3 seconds, respectively). III. CONTINUOUS MONITORING AND DETECTION APPROACH The physical partitioning and form-factor of the system have important implications to patient usability, power consumption, and robustness. For instance, EEG sensing must be distributed around the scalp in order to acquire spatial channels, but SVM classification (over the multidimensional feature vector) must be centralized. As a result, the intermediate instrumentation, computation, and communication tradeoffs determine the ap- propriate system topology. Further, a critical application con- sideration is that, for chronic seizure detection, no cables can Fig. 1. 18-channel EEG showing onset of patient seizure (ictal); electrical onset originate from the scalp, since these pose a strangulation hazard occurs 7.5 sec before the clinical onset, which is characterized by muscle re- in the case where the patient begins convulsing. Accordingly, flexes causing the large excursion artifacts. some form of wireless transmission from the scalp is essential. For sensing robustness, the acquisition circuitry (e.g., instru- mentation amplifier and ADC) is kept as close to the electrodes as possible to mitigate EMI and mechanical disturbance on wires carrying the microvolt EEG signals. As a result, the in- strumentation amplifier (and, to a lesser extent, also the ADC) must be distributed along with each electrode. The digitized EEG recordings can be robustly transmitted (i.e., wireless EEG) for central processing. However, local pro- cessing is beneficial for minimizing communication cost. In Table I, the system power for wireless EEG (where both fea- Fig. 2. Seizure detection algorithm employing spectral analysis feature extrac- ture vector extraction and SVM classification are performed re- tion and SVM classification. motely) is compared to that with local processing (where feature vector extraction is performed locally and only SVM classifi- basis, thereby simultaneously improving sensitivity and speci- cation is performed remotely). The power numbers are based ficity of detection. The following subsection briefly describes on actual measurements of the hardware prototype assuming the approach and parameters used in this system. 18 EEG channels. The radio used is a commercially available low-power transmitter, ChipCon CC2550 [11], and its power A. Seizure Detection Algorithm consumption is for duty-cycled operation at the required data- Fig. 2 illustrates the detection algorithm. First, the EEG chan- rates (including idle, start-up, and active transmission modes). nels are processed to extract specific bio-markers that are rele- By performing local processing to extract the feature vectors for vant for seizure detection. Clinical studies have determined that transmission, the radio data-rate is reduced by a factor of over 40 seizure onset information is contained in the spectral energy dis- compared to complete wireless EEG transmission. Through this tribution of the patient’s EEG [9]. Accordingly, in this SoC the computation-versus-communication tradeoff, local processing spectral energy of each channel is extracted to seven frequency reduces the total system power on the scalp by a factor of 14 for bins over a two second window in order to form a feature vector. the radio considered.