F Onorati1, G Regalia1, C Caborni1, R Picard1,2. 1 F Onorati1, G Regalia1, C Caborni1, R Picard1,2* 1. Empatica, Inc, Cambridge, MA and Milan, Italy, www.empatica.com 2. MIT Media Lab, Massachusetts Institute of Technology, Cambridge, MA *emails: fo@empatica.com or picard@media.mit.edu Improvement of a convulsive seizure detector relying on accelerometer and electrodermal activity collected continuously by a wristband Rationale Acceleration (ACM), measured through a wrist-band, has been used to automatically identify the motion of convulsive seizures such as generalized tonic-clonic seizures (GTCSs) [1] [2]. However, ACM-based seizure detectors have high false alarm rates, which might prevent daily life use. Electrodermal activity (EDA) is a physiological signal reflecting the activity of the sweat glands driven by the Sympathetic Nervous System. Direct stimulation of some subcortical regions [3] elicits ipsilateral EDA responses. Using both EDA and ACM features improves the classification performance of a GTCS detector, reducing the false alarm rate [4] (16 GTCSs from 7 patients). Using a larger feature set further lowers the false alarm rate while preserving high sensitivity [5] (20 GTCSs from 9 patients). Here we present the results of further improvements, using a smaller set of features, better matched to use in patient-friendly wearable devices: 38 GTCSs from 18 patients, obtaining similarly high sensitivity (92-95%) with low false alarm rates (0.56-2.26 per day). Methods Labeled seizure data were collected during clinical video EEG (v-EEG) monitoring. Data: 44 recordings from 18 patients wearing a wrist-band able to record EDA and 3-axes ACM (Figure 1(a)). EDA and ACM signals were analyzed off-line using proprietary software to clean the data and extract signal features over a 10 sec. window every 2.5 sec (overlap: 75%). Support Vector Machine (SVM) classifiers were trained: one with 46 features (SVM_46), one with 30 features (SVM_30). A leave-one-patient-out cross-validation approach was used to test classifier sensitivity (Se) and false alarm rate (FAR), defined as number of false alarms per day. The optimal decision threshold was selected by receiver operating characteristic (ROC) curve analysis. (a) (b) PATIENT 1 PATIENT 2 FEATURE EXTRACTION GTCS DETECTOR ACM EDA FEATURE SET NO GTCS GTCS (ALARM) Figure 1. (a) EDA and 3-axes ACM signals of two patients recorded during a generalized tonic-clonic seizure (GTCS) with a wrist-worn device. (b) Schematic workflow of the GTCSs detector. Results Recordings included 38 GTCSs from 18 patients over a total of 1027 hours (42.7 days). Both classifiers show acceptable FAR while keeping Se higher than 90% (Figure 2). SVM_46 at Se=92% (i.e., 35 seizures detected out of 38) showed FAR=0.56 and at Se=95% (i.e., 36 out of 38) FAR=2.26. SVM_30 achieved similar performance (Se=92% and FAR=0.74, Se=95% and FAR=2.02), using fewer features. Conclusions A seizure detection system based on ACM and EDA features was developed using clinical data collected from a larger number of patients and seizures with respect to previous work, capturing greater subject variability of GTCS expression. The classifier we obtained allows a higher seizure detection rate while maintaining an acceptable false alarm rate. Furthermore, it is efficiently integrated into a wearable wristband to provide real-time alarms of ongoing seizures. When using SVM_30 with a threshold set to provide Se 95%, most of the patients (10/18 = 55%) had less than 1 false alarm every 2 days and most (16/18 = 88%) had ALL of their GTCSs detected (Figure 3). References [1] J Lockman et al., “Detection of seizure-like movements using a wrist accelerometer”, Epilepsy and Behavior, v. 20, no. 4, pp. 638-641, 2011. [2] S. Beniczky et al., “Detection of generalized tonic–clonic seizures by a wireless wrist accelerometer: A prospective, multicenter study”, Epilepsia, v. 54, no. 4, pp. e58–e61, 2013. [3] C. A. Mangina and J H Beuzeron-Mangina, “Direct electrical stimulation of specific human brain structures and bilateral electrodermal activity”, Int J of Psychophysiology, v. 22, no. 1–2, pp. 1–8, 1996. [4] M-Z. Poh et al., “Convulsive seizure detection using a wrist-worn electrodermal activity and accelerometry biosensor,” Epilepsia, v. 53, no. 5, pp. e93–e97, 2012. [5] G. Regalia et al., “An improved wrist-worn convulsive seizure detector based on accelerometry and electrodermal activity sensors”, American Epilepsy Society annual meeting 2015, abs no. 3096, 2015. Figure 3. Distribution of FAR (top) and Sensitivity (bottom) of the new SVM_30 GTCS detector across data from 18 patients (38 GTCSs, 1027 hours of recordings). Figure 2. Performance of the SVM classifier based on 46 EDA and ACM features (SVM_46, blue line) and re-trained by using 30 features (SVM_30, red line). With gratitude to Drs. Dan Friedman, Tobias Loddenkemper, Claus Reinsberger, Rani Sarkis, and their teams at NYU, Boston Childrens Hospital, and Brigham & Womens Hospital, Boston and to Jonathan Bidwell and his teams at Emory and Childrens Healthcare of Atlanta Hospital, Atlanta, for their generous support collecting sensor data and providing gold standard v-EEG labeled data.