Aims: Deep learning has emerged as a powerful tool for predicting future health outcomes directly from nocturnal polysomnography (PSG) data. Prior work has primarily emphasized temporal correlations between modalities, often at the expense of preserving modality-intrinsic information. This study introduces a multimodal residual spectrogram encoder designed to disentangle private modality-specific features from shared representations for predicting cognitive impairment from PSG.
Methods: We used multimodal PSG recordings from the George B. Moody PhysioNet Challenge 2026 dataset to learn patient-level representations for predicting cognitive impairment within 3 to 7 years. The model was first pretrained using a contrastive predictive coding (CPC) framework on the SleepFM dataset and subsequently fine-tuned for the cognitive impairment prediction task. Each PSG channel was standardized using per-recording statistics, after which 5-second intervals were transformed into log-spectrograms via the Short-Time Fourier Transform (STFT). These spectrograms were processed by convolutional residual blocks to capture frequency-domain patterns. A modality adapter then generated group-specific representations for the four modalities: brain (EEG), respiratory, EMG, and ECG. Modality interactions were modeled through intermediate encoder stages, followed by fusion of the embeddings to form a shared representation branch. An orthogonality loss encouraged disentanglement between shared and private embeddings. For final prediction, mean-pooled features from the full recording were passed through a linear classifier and trained with binary cross-entropy (BCE) loss.
Results: In 5-fold cross-validation on a random subset of the development data (N=410), the model achieved a mean AUROC of 0.59 (min = 0.51, max = 0.71). On the hidden validation set of the challenge, it obtained an unofficial score of 0.574 (Team Dachshund Legacy).
Conclusion: The results demonstrate feasibility of a modality adaptive resblock architecture, though further improvements are required to reach clinical utility.