Biomedical Research Lead Columbia University Littleton, Massachusetts
Rationale: Currently, the cycle of care for epilepsy is hamstrung by a lack of quantitative solutions for monitoring patients outside the clinic. Physicians often rely upon inaccurate, self-reported seizure counts entered in electronic diaries to inform clinical decisions. This problem is exacerbated for patients that exhibit focal onset with impaired awareness (FIA) seizures. Due to their non-convulsive presentation and associated patient loss of consciousness, FIA seizures are far less likely to be recorded in eDiaries than other subtypes [1]. In this research, we aim to develop deep learning models that use EEG to perform accurate, patient-independent FIA seizure detection. While research has been focused on improving convulsive seizure detection by utilizing various signal processing approaches, limited work has been performed to optimize FIA seizure detection models. In addition, few studies have interrogated whether information contained in other seizure subtypes can help deep learning models achieve better FIA seizure detection performance through transfer learning. Methods: EEG data from the Temple University Hospital Seizure Detection (TUSZ) Corpus, acquired through a standard 10-20 montage, was utilized in this analysis [2]. Data was split into ten-second epochs. All epochs were then average referenced. After preprocessing, the signals were converted into four modalities to develop classification models: time series, temporal FFT, short-time fourier transform (STFT)-based spectrograms, and morlet continuous wavelet transform (CWT)-based scaleograms. ResNet50 was utilized as the base architecture. Baseline models were developed on each signal modality using a 33 patient training set containing a balanced number of FIA seizure and non-seizure (NS) epochs. Next, EEG epochs containing eight seizure types, including FIA, and a balanced number of NS epochs were utilized to pretrain each of the four models. Fine tuning was then performed on the original training set. Each model was evaluated on a separate, balanced seven patient test set. Results: Baseline time domain-based models outperformed STFT, temporal FFT, and CWT models (Figure 1). Pretraining improved model performance for each modality, except CWT (Figure 1). The top-performing pretrained model, also based on time-series, achieved an 86% test set accuracy (chance = 50%), an AUC of 0.94, a recall of 90%, and a precision of 83% (Figure 2). This model robustly detected each test set patient’s FIA seizures, with a detection accuracy of at least 78% per individual. Conclusions: This research shows that highly competitive EEG-based FIA seizure detection algorithms can be developed by pretraining models on a diversity of seizure subtypes. Future work should interrogate the potential benefits of cross-subtype pretraining for the detection of other subtypes.
[1] Hoppe C. et al. Epilepsy: accuracy of patient seizure counts. Arch Neurol. 2007;64(11):1595–9. [2] Shah, V. et al. The Temple University Hospital Seizure Detection Corpus. Front Neuroinform. 2018; 12: 83. Funding: Please list any funding that was received in support of this abstract.: N/A
