Aim: The study aims to assess the applicability of machine learning approaches to predict cognitive impairment (CI) from overnight polysomnography recordings, as an entry to the George B. Moody PhysioNet Challenge 2026. Methods: Our initial approach used Light Gradient-Boosting Machine (LightGBM) on a dataset of 622 labeled subjects (311 CI+, 311 CI-) drawn from three acquisition sites with heterogenous recording hardware, montages, and sampling rates. Feature extraction comprised four blocks: demographics [9 dimensions], physiological [70 dimensions], algorithmic [15 dimensions], and human annotations [12 dimensions], totaling 106 dimensions per subject. Epoch-level Hjorth parameters (activity, mobility, and complexity) were computed across seven signal types (EEG, EOG, chin EMG, leg EMG, ECG, respiration, and SpO2) plus per-stage chin EMG features for REM, N2, and N3 separately. Additionally, features like sleep architecture, event densities (apnea-hypopnea index [AHI], arousal index, limb movement index), and fragmentation metrics were derived from expert-scored staging (human annotations block) or automated annotations (algorithmic block, which also included CAISR classifier probabilities). Cross-site signal harmonization was handled through per-site z-score normalization estimated from a held-out supplementary set, addressing ~40x gain variation in leg EMG and unit discrepancies in SpO2 across sites. The LightGBM classifier was trained using leave-one-site-out cross-validation; the number of boosting rounds was calibrated to the largest site's optimal iteration to prevent overfitting on smaller folds. Results: Our unofficial phase submission (ID 1326) yielded AUROC:0.588, AUPRC:0.114, accuracy: 0.867, and F-measure:0.115. Univariate Mann-Whitney analysis identified 27 features with significant CI separation (p<0.05). Chin EMG activity dominated (particularly during REM sleep with rank-biserial r=-0.22, p<0.001), with CI+ subjects showing lower EMG amplitudes across all stages, consistent with neurodegeneration-associated REM atonia changes. EEG mobility and CAISR arousal probability also contributed meaningfully, while demographic features were largely uninformative. Conclusion: Machine learning approaches may be useful for identifying polysomnography features indicative of cognitive impairment.