Improving wearable-based seizure prediction by feature fusion using an explainable growing network.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning in Healthcare

Background:

• Epilepsy seizure unpredictability poses significant burdens on patients and caregivers, impacting health and quality of life.
• Wearable devices offer a promising avenue for non-stigmatizing, user-friendly seizure prediction using physiological data.

Purpose of the Study:

• To develop and evaluate a patient-agnostic seizure prediction method utilizing group-level patterns from multi-patient data.
• To enhance seizure prediction accuracy through the fusion of multiple machine learning techniques and incremental learning.

Main Methods:

• Employed a supervised long-short-term network (LSTM) combined with an unsupervised deep canonically correlated autoencoder (DCCAE) and time-of-day pattern analysis.
• Fused features from these models using a growing neural network for incremental learning.
• Utilized Shapley additive explanations (SHAP) to analyze feature contributions and assessed the impact of data quality and clinical variables.

Main Results:

• The proposed method achieved an average prediction accuracy of 81.7%, an improvement of 7.3% over the baseline LSTM (74.4%).
• The method outperformed the LSTM in 84% of patients, with the growing network fusion improving accuracy by 9.5% compared to all-at-once fusion.
• Analysis revealed the influence of preictal data duration, wearable data quality, and clinical variables on prediction performance.

Conclusions:

• A patient-agnostic seizure prediction method using wearable devices can significantly improve prediction accuracy.
• The fusion of LSTM, DCCAE, and temporal patterns via a growing neural network offers a robust approach for epilepsy management.
• This technology has the potential to enhance patient safety and quality of life for individuals with epilepsy.