Overnight polysomnography captures neurocognitive decline through subtle disruptions of sleep microstructure, but no single index reliably discriminates cognitively impaired (CI) from unimpaired individuals. We present a validated pipeline that detects CI from automated CAISR sleep annotations using a compact, interpretable feature set and a site-normalised penalised logistic regression. Forty-two features were extracted per participant: standard sleep architecture (AHI, arousal index, PLM index, WASO, SOL, REM latency, stage percentages), temporal complexity (Shannon entropy, Lempel-Ziv complexity, cross-transition entropy, N3/REM bout statistics, hypnogram distribution asymmetry), and CAISR probability channel means. Three engineering corrections were applied: (1) respiratory events were segmented as continuous physiological episodes—collapsing multi-label classifier flicker and enforcing AASM ≥10 s duration—using TST as the AHI denominator; (2) stage gap-filling was limited to two epochs, assigning Wake to longer artefact spans; (3) epoch stages were sampled from the median of the central 20 samples per epoch. Per-site Z-score normalisation was embedded in the pipeline to remove centre effects. Permutation importance under 3-fold site-stratified cross-validation identified the top-8 features; ElasticNet hyperparameters were tuned via nested cross-validation. The submitted model achieved an official challenge AUC of 0.583 (Entry #1220, Team apple_dream, Rank 76). In the training set, the dominant predictor was arousal index (ΔAUC +0.096), which was paradoxically lower in CI patients (p < 0.006, Mann-Whitney) across all three AHI severity strata and after residualising for AHI—consistent with impaired cortical arousal responsiveness. Secondary contributors in the training set were PLM index, early-night REM proportion, hypnogram entropy, N3-to-Wake transitions, and WASO. Notably, the paradoxically reduced arousal index in CI patients—despite equivalent apnoea burden—suggests a failure of cortical arousal responsiveness rather than a protective adaptation, potentially reflecting cholinergic and noradrenergic deficits that are hallmarks of neurocognitive decline. This multivariate pattern is robust to site effects and to the severe apnoea prevalent in this cohort..