Predicting Cognitive Impairment from Polysomnography with Tabular Foundation Model and Markov Transition Matrix

Kelvin K Nguyen, Arshia Ilaty, Andy Y Smithwick, Arash Shokouhmand, Kingson Zhang, Elahe Khatibi, Sogol Sanjaripour, Jungsoo Lee, Shadi Manafi, shahab nikkhoo, Zaniar Ardalan, Saman Parvaneh
Edwards Lifesciences


Abstract

Cognitive impairment develops slowly, and overnight sleep changes years before a diagnosis is recorded. The George B. Moody PhysioNet Challenge 2026 asked participants to predict, from a single overnight polysomnogram (PSG), whether a patient later receives a diagnosis of mild cognitive impairment, Alzheimer's disease, or dementia. Our team (SleeperAgents) built a feature-based pipeline that reads each recording and its automated annotation file once and turns them into 218 features in nine blocks, covering demographics, whole-night time-domain statistics, annotation summaries, heart rate variability, band power and respiratory descriptors, stage-conditional cross-modal measures, first-order Markov sleep-stage dynamics, and stage-aware EEG biomarkers. A class conditional Markov log-likelihood ratio, cross-fit across recording sites, is appended as one extra column. An XGBoost classifier and a TabPFN v2 tabular foundation model are then trained on the resulting 219-column matrix, each calibrated with isotonic regression on leave-one-site-out out-of-fold predictions and blended with a weight tuned on the same predictions. Our entry received an age-conditioned AUROC of 0.604 on the official hidden validation set.