Accurate prediction of future cognitive impairment can support earlier identification of at-risk individuals and enable trials of novel interventions for conditions such as Alzheimer's disease. For the 2026 PhysioNet Challenge, our team (Headington Dreamers) used a combination of hand-crafted features and a lightweight neural network to address this task.
Inspired by the literature, we computed 10 features from the C3 and C4 EEG channels for each 30-second sleep epoch, including band-limited kurtosis and measures of signal entropy. These were combined with one-hot encoded sleep stages to produce a 15D feature vector for each sleep epoch. We then trained a lightweight, gated attention pooling model which combined features from 128 randomly sampled sleep epochs to generate a prediction for each patient. An ensemble of five models was trained using 5-fold cross validation to minimise the binary cross-entropy loss against cognitive impairment labels. Using the ensemble average output, we obtained an official score (AUROC) of 0.739 on the held-out set, and a nested cross-validation AUROC of 0.721 on the training set. This approach outperformed a logistic regression baseline (official score of 0.647) which used night-level summary features such as total deep sleep and sleep spindle density.
For the official phase, our hand-crafted features will be replaced with embeddings from a pre-trained foundation model that we are developing. After self-supervised learning using next-token prediction on datasets including the Sleep Heart Health Study (SHHS), our foundation model provides highly effective low-dimensional embeddings of physiological data. For example, a linear classifier trained on embeddings from an early version of our foundation model obtains expert-level sleep staging accuracy on SHHS. By automatically extracting features from additional modalities, such as the F1 and F2 EEG channels, and by also capturing cross-modal interactions, we expect this model will help to further improve predictive performance.