Multimodal Relapse Prediction

Overview

Relapse episodes are rare, labels are scarce, and patients differ widely, so models trained on one group often fail on the next. This project uses the 2nd e-Prevention challenge data: nine patients, five smartwatch sensors, and about 3,300 patient-days. Every model is evaluated with leave-one-patient-out cross-validation, so each fold tests on a patient the model has never seen.

What I built

• Built the supervised sequence transformer: a transformer encoder over 7-day windows, with left-padding and attention masking so every test day gets a prediction.
• Handled class imbalance with SMOTE on flattened training windows, focal loss, and label smoothing.
• Ran a global hyperparameter search across all folds and an ablation over loss functions, weight decay, stochastic depth, and positional encoding.
• Stabilized the focal-loss and label-smoothing configuration with AdamW and gradient clipping, which gave the best result.