Weak self-supervised learning for seizure forecasting: a feasibility study.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Biomedical Engineering

Background:

• Time-series data exhibit autocorrelation, which can be leveraged for predictive modeling.
• Manual labeling of physiological signal recordings for patient monitoring is time-consuming and burdensome.
• Personalized forecasting models improve performance in medical applications like seizure prediction.

Purpose of the Study:

• To propose an artificial intelligence system for continuous improvement in event prediction using self-supervised learning on unlabelled data.
• To reduce the need for manual data annotation in medical patient monitoring.
• To develop and validate individualized forecasting models for seizure prediction.

Main Methods:

• Utilizing a detection model to generate weak labels on-the-fly.
• Training a prediction model on a time-shifted input data stream using these generated labels.
• Applying a self-supervised approach to train individualized forecasting models on patient-specific neurophysiological data.

Main Results:

• Demonstrated feasibility of a self-supervised approach for personalized forecasting models.
• Achieved an average relative improvement in sensitivity of 14.30% for early seizure forecasting.
• Reduced false alarms by 19.61% in early seizure forecasting across 10 patients.

Conclusions:

• Self-supervised learning effectively harnesses autocorrelation in time-series data for event prediction.
• The proposed system enables personalized forecasting models without extensive manual labeling, crucial for medical monitoring.
• This proof-of-concept paves the way for low-power neuromorphic neuromodulation systems for continuous patient monitoring.