Abstracts

Rationale: Measurement of wrist acceleration (ACM) by means of wearable devices has been exploited to automatically detect ongoing motor seizures in patients with epilepsy (Epilepsy and Behav 2011, 20, 638-641; Epilepsia 2013, 54(4), e58-e61). Nevertheless, such seizure detectors can show high false alarm rates in active patients, which might hinder their use in daily life. Electrodermal activity (EDA) is a physiological signal reflecting sympathetic activity. Large ipsilateral EDA responses are elicited via direct stimulation of several subcortical regions (Int J Psychopysiol 1996, 22, 1-8). Measuring a combination of EDA and ACM has been previously shown to enhance specificity, i.e. to reduce the false alarm rate, in detection of secondary generalized tonic-clonic seizures (GTCS) (Epilepsia 2012, 53(5), 93-7). Nevertheless, the aforementioned approach requires further improvements in generalization capability and in further reducing false alarm rate for use in the widest variety of daily activities. Accordingly, in this contribution we report the performances of four ACM and EDA-based seizure detectors fed with different feature sets, trained on a higher number of seizures than in our previous work (Epilepsia 2012, 53(5), 93-7).Methods: Data were collected during clinical video EEG monitoring, and consist of 31 recordings taken from 9 patients wearing a wrist device recording EDA and three-axis ACM signals. EDA and ACM recordings were analyzed off-line using proprietary software to clean the data and extract four different feature sets. The properties of each feature set are summarized in Table I. Each set of features was used to train a Support Vector Machine (SVM) classifier. A leave-one-patient-out approach was employed in order to evaluate the sensitivity (Se) and the false alarm rate (FAR) of the classifier. The model was further improved by maximizing the cost function of a receiver operating characteristic (ROC) curve.Results: The 9 patients’ recordings included 20 GTCS over a total of 738 hours. For all feature sets, the classifier was able to detect 19 out of 20 seizures (Se: 95%). Feature sets 2, 3 and 4 show a significantly lower FAR with respect to Feature set 1, previously published (Epilepsia 2012, 53(5), 93-7). We reduced the FAR from 2.02 to 0.48 using feature set 3 instead of the original features, i.e. an overall reduction by a factor of 4.2. Figure 1 depicts the FAR achieved by the 4 classifiers.Conclusions: A seizure detection system was improved based on combining ACM and EDA information. The use of the new features preserved the current high Se and significantly reduced the FAR to less than one false alarm every 48 hours. After having selected the final feature set, the classifier can be integrated in a hardware platform to provide reliable real-time seizure alarms for caregivers.