Cross-Site Evaluation of Demographic and Sleep-Annotation Features from Polysomnography for Cognitive Impairment Prediction

Aans Noufal1, Emadaddin Al-Qadasi1, Florian Finn Keuren2, Marc Volpe-Prignano Garcia1, Mykhailo Bratskov3
1Eindhoven University of Technology, 2TU Eindhoven, 3Technical University of Eindhoven


Abstract

Polysomnography (PSG) captures multimodal physiological signals during sleep that may reflect early markers of neurological dysfunction and future cognitive impairment. However, the generalizability of commonly used PSG-derived summary features across clinical sites remains unclear. We evaluate this question using demographic and sleep-annotation features in a cross-site binary classification task.

We constructed a 35-dimensional feature representation combining demographic variables (age, sex, race, body mass index) with sleep architecture metrics, including stage proportions (Wake, N1, N2, N3, REM), sleep efficiency, respiratory indices (apnea-hypopnea index), arousal and limb movement indices, REM-specific event counts, and summary statistics of event durations. Features were derived from automated Complete AI Sleep Report (CAISR) PSG annotations. Missing annotation channels across recordings were handled via NaN-padding, preserving feature dimensionality across heterogeneous sites. A gradient-boosted decision tree model (XGBoost, maximum depth 3, binary logistic objective) was trained for classification.

To evaluate robustness under dataset shift, we performed leave-one-site-out cross-validation across three clinical sites (S0001, I0002, I0006). The model achieved a mean AUROC of 0.551 (±0.031), with per-site AUROCs of 0.562, 0.507, and 0.562. The official submission score was 0.404, with post hoc analysis identifying a label encoding error affecting submitted predictions; corrected cross-validation results are reported above. Performance remained close to chance despite inclusion of clinically motivated sleep architecture and physiological summary features.

These findings indicate that standard PSG-derived summary features and demographic variables provide limited predictive value for cross-site generalization in cognitive impairment risk prediction. The results further suggest that improving performance likely requires more expressive representations of raw physiological signals, including EEG, ECG-derived heart rate variability, and respiratory waveform morphology, to capture subtler temporal and spectral dynamics not represented in summary annotations.

Team Neurion, Eindhoven University of Technology.