We aim to predict whether a patient will receive a cognitive impairment diagnosis 3 to 7 years after a sleep study using full polysomnogram features, including EEG, ECG, EMG, EOG, respiratory measures, and demographics. Besides the competition challenge, there is a substantial challenge in using this breadth of data, including missing signals, time series data vs. demographic data, and varying sampling and annotation frequencies. The size of the dataset is also a challenge, as loading the data takes a significant amount of time and cannot all reside in memory concurrently.
We combine feature embeddings from SleepFM and CAISR with embeddings generated from our proposed embedding model. Our model uses a channel-level transformer with temporal encodings so that channels with different sampling rates can be processed. This transformer is pre-trained to predict sleep stage labels, after which class-token embeddings are cached to form a uniformly sampled low-frequency sequence across the full night. Embeddings are then pooled across channels using an attention mechanism resilient to missing data.
We concatenate our novel embeddings with those of SleepFM and CAISR, along with engineered features such as heart-rate variability, spectral power, spindle morphology, and REM EMG activity, together with whole-night features such as sleep-stage proportions, total wake time, spindle counts, and demographic variables. These learned and engineered representations are processed in a final classifier to predict future cognitive impairment. To support long recordings and reduce memory demands, training uses streaming, on-demand window generation instead of precomputing all windows in memory.
The Team Blue and Gold unofficial leaderboard AUROC score is 0.569, which represents the results of the full-night engineered feature component of our approach. Ten-fold cross-validation using only SleepFM embeddings yielded an AUROC of 0.597. Results from ensembling all embeddings are forthcoming, and results from individual embeddings will be provided in an ablation study.