Age-Residualized Physiological Biomarkers and Gradient-Boosted Trees for Predicting Cognitive Impairment from Polysomnography

Paul Fischer1, Simon Böhi1, Sidaty El Hadramy1, Volodimir Buchakchiyskiy2, Hélène Corbaz3, Max Krähenmann1, Philippe C. Cattin1, Ece Özkan Elsen4
1University of Basel, 2University of Base, 3University of Basel, University Hospital of Basel, 4University of Basel


Abstract

We present our entry, submitted by team Koalalition, to "Screening for Cognitive Impairment During Sleep Studies: The George B. Moody PhysioNet Challenge 2026". Rather than learning directly from raw multi-channel signals, we extract a bank of 136 literature-motivated physiological features from the measurements, stratified by sleep stage. We further compute an age-residual feature for a curated set of 23 biomarkers with established age trends. The resulting tabular representation is classified with a bagging ensemble of 15 gradient-boosted trees (XGBoost), with hyperparameters selected by randomized search under stratified cross-validation. On the official challenge validation set, our submitted model trained on the large version of the dataset reaches an age-conditioned AUROC of 0.669, ranking in place 136.