Abstracts

Rationale: Proof-of-principle studies suggest that seizure prediction is achievable with wearable devices that record physiological signals such as electrodermal activity (EDA), heart rate (HR), and body temperature (TEMP). Feasibility was shown with different machine learning techniques, such as long short-term memory networks (LSTM) [1]. In this study, we develop a novel seizure prediction method from wristband data that combines an unsupervised deep canonically correlated autoencoder (DCCAE) and a supervised LSTM. We hypothesize that DCCAE and LSTM focus on different types of biomarkers and complementary information, and their combination could improve prediction accuracy. [1] Meisel et al., “Machine learning from wristband sensor data for wearable, noninvasive seizure forecasting,” Epilepsia, 2020


Methods: Out of 450 patients within the Detect, Predict, and Prevent Seizures study at Boston Children’s Hospital from 2015 to 2021, we included 38 patients diagnosed with epilepsy with generalized tonic-clonic (GTC) or focal to bilateral tonic-clonic (FBTC) seizures. All patients were enrolled in the long-term video-EEG monitoring unit and wore a wearable biosensor (Empatica® E4, Milan, Italy) on their wrist and/or ankle that recorded HR, EDA, and TEMP. We determined seizure onset by reviewing video EEG data and adding a buffer of 30 seconds before to account for possible synchronization inaccuracies. We cut 5 minutes of data before the seizure onset and defined it as preictal. The interictal data consisted of 5 continuous minutes within the interictal phase, defined as two hours before seizure onset to the maximum available duration without any previous seizure occurrence for the same patient. These 5 minutes were either randomly extracted from the interictal phase or were chosen to be maximally far away from seizure onset. All 5-minute segments were cut into 15-second windows with no overlap. Using all modalities, a supervised LSTM neural network was optimized for feature selection to classify preictal and interictal data. In parallel, a DCCAE was trained in an unsupervised fashion for extracting highly correlated features from EDA and HR. The extracted features were then concatenated to train a final classification model, as shown in Fig. 1. We reported classifier performance for 10-fold cross-validation and used the evaluation split for early stopping.


Results: We included 38 patients (47% female; median age: 14 years) with 59 seizures (13 GTCs and 46 FBTCs). Table 1 shows results for randomly chosen interictal and maximally far away interictal segments. Our proposed method outperforms both DCCAE and the LSTM network in both cases.


Conclusions: We combined unsupervised features learned using DCCAE from wearable recordings with a supervised LSTM to accurately predict seizures in pediatric patients with epilepsy. Utilizing this complementary information improves seizure prediction accuracy.


Funding: This study was supported by the Paderborn University research award and the Epilepsy Research Fund.