Abstracts

Rationale:
Tracking seizures is crucial for epilepsy monitoring and treatment evaluation. While EEG remains the gold standard for seizure detection, non-invasive, easy to use wearable devices may be better tolerated and more suitable for long-term ambulatory monitoring. While wearable devices have been shown to detect generalized tonic and tonic-clonic seizures, detection of other seizure types is unclear. Here exists a critical need for reliable, automatic, non-intrusive, methods to detect broader seizure types. We aimed to evaluate seizure detection performance across a broad spectrum of epileptic seizures by means of a wrist-worn device.
Method:
We enrolled Boston Children’s Hospital’s Epilepsy Monitoring Unit patients between 2015 and 2017. Wearable sensors (E4, Empatica, Milan, Italy), worn on patients’ wrist or ankle, recorded body temperature (TEMP), electrodermal activity (EDA), accelerometry (ACC), and photoplethysmography (PPG) to provide blood volume pulse (BVP) from which heart rate (HR) can be derived. We used EEG seizure onset and offset as determined by a board certified epileptologist. The dataset was cleaned, and low-quality data segments caused by switched-off wristbands or device/battery failures were removed. The final dataset used for this study contains 192 Patients (50% female, median age 10.01 years) and 722 epileptic seizures. Wristband and EEG clocks were synchronized at the recording start. Timing drift between the clocks was measured (Fig. 1a,b) and allowed to derive a three-step compensation scheme: i) a 13 seconds off-line timing drift compensation was added to the wristband data, ii) 20 seconds pre-seizure and 20 seconds post-seizure end windows were added , iii) seizure annotations were added to the wristband data.  We applied a 5-fold patient-wise cross-validation scheme to BVP, HR, EDA, ACC, and TEMP data (Fig. 1c). Train and test sets contained different groups of patients for all model runs. Random under-sampling was used to balance the training data. Model performances were evaluated on the full test data. Five Machine Learning methods (KNN, XGBoost, CNN, MLP, LSTM) were applied to raw time series sensor data. As post-fusion analysis, the probabilistic distributions from the individual models were combined to infer class labels and post-processing techniques using a moving average were investigated.
Results:
Best results are shown in Table 1. All sensor modalities perform better than chance, and data fusion methods outperform single modalities with an AUC-ROC of 66.35%.
Conclusion:
Automatic epileptic seizure detection using machine learning and wearables data is feasible. Preliminary results show better-than-chance seizure detection across a broad range of epileptic seizures. Future improvements may consider clinical chrono-epileptological variables, such as seizure duration, semiology, and etiology or syndrome, and alternative data balancing, pre- and post-processing, fusion and ensemble learning methods. Thus, while findings suggest feasibility, performance following future adjustments may improve further.
Funding:
:N/A