Predicting Cognitive Impairment from Polysomnography Using Hybrid Hand-Crafted and Deep Learning Features with Gradient Boosting

Efrain Manuel Magana and Juan Ignacio Ortega Flores
Pontificia Universidad Católica de Chile


Abstract

Early identification of individuals at risk for cognitive decline represents a critical unmet need in clinical neurology. The George B. Moody PhysioNet Challenge 2026 tasks participants with predicting future diagnosis of mild cognitive impairment, Alzheimer's disease, or dementia from a single overnight polysomnography (PSG) recording and basic demographic data, scored by AUROC between confirmed cases and long-term cognitively intact controls. We developed a multi-stream feature extraction pipeline applied to the 622-record training set from the Human Sleep Project (three institutions: BIDMC, Emory, Stanford). Signals across ten channel groups (frontal, central, and occipital EEG; EOG; chin and leg EMG; ECG; respiratory flow and effort; and SpO2) were resampled to 128 Hz and segmented into 60-second epochs. Four hand-crafted feature sets were computed: spectral features (band power, relative power ratios, spectral entropy, and aperiodic slope across delta, theta, alpha, sigma, and beta bands); epoch statistics (mean, variance, trend, percentiles, and IQR); cross-channel coherence across clinically motivated channel pairs; and amplitude variability metrics. Instead of using raw demographic and annotation variables directly, we constructed interaction terms such as age x N3 deficit, BMI x AHI, and age x arousal index, then pruned the combined set iteratively based on cross-fold importance variance. In parallel, a deep learning encoder was trained on sequences of per-epoch descriptors spanning the full night, with epochs sampled for uniform temporal coverage. Penultimate-layer embeddings were extracted using out-of-fold inference to avoid leakage, concatenated with the pruned hand-crafted features, and passed to a regularised LightGBM classifier. Internal 10-fold stratified cross-validation yielded an AUROC of 0.900 using hand-crafted features alone, improving to 0.930 +/- 0.039 after adding deep learning embeddings. Official validation results are pending. Ongoing work focuses on additional deep feature streams and calibration under the realistic 5-15% prevalence of the validation and test sets.