Aims: Cognitive impairment (CI) refers to difficulties in memory, language, thinking, or judgment that go beyond typical aging. Neuropsychological tests, considered the gold standard, facilitate early detection but are not routinely recommended for asymptomatic individuals. Physiological data may serve as an objective source for prediction, yet their potential remains underexplored. In this study, we investigate the use of selected polysomnography (PSG) signals for early prediction of CI.
Methods: We analyzed PSG recordings (N=622, 264 females, mean age 69.9 +/- 8.6 years) from the Human Sleep Project dataset. Our initial approach employed feature extraction across demographics and multiple physiological signals (EEG, ECG, SpO2) followed by feature selection and classification using an XGBoost ensemble with 5-fold cross-validation. We systematically evaluated the contribution of different signal modalities, with a particular focus on EEG-derived features.
Results: The baseline model achieved an ROC-AUC of 0.493 for the challenge validation set. Incorporating sleep stage specific heart rate variability and oxygen saturation features alongside model regularization parameters improved performance to 0.515 and 0.502. The inclusion of EEG band power features resulted in a substantial performance gain, increasing ROC-AUC to 0.653 for the challenge validation set despite the low cross-validation score of 0.579 of the training set. These findings highlight the strong predictive value of EEG in detecting CI from sleep studies.
Conclusion: EEG-derived features provide the most informative signal for CI prediction in PSG data. The next project phase will target major performance gains by training improved deep learning models on spectrograms of raw signals and leveraging transfer learning from the full Human Sleep Project and Sleep Heart Health Study datasets.