Early identification of cognitive impairment from routine clinical recordings could enable timely intervention. Polysomnography (PSG) captures electrophysiological dynamics during sleep that may reflect neurodegenerative processes years before clinical diagnosis. We, team CardioAI, present a method that predicts future cognitive impairment exclusively from PSG signal characteristics and automated sleep annotations, without relying on demographic features such as age, sex, race, or BMI. We extracted multichannel PSG signals from training data spanning three clinical sites (780 recordings). Channels were standardized via a rule-based renaming pipeline and bipolar re-referencing. One representative channel per signal group (EEG, EOG, chin EMG, leg EMG, ECG, respiratory, SpO2) was selected, resampled to 64 Hz, and truncated to one hour. For each channel, we computed 20 features: eight time-domain (standard deviation, mean absolute value, RMS, zero-crossing rate, skewness, kurtosis, Hjorth mobility and complexity), ten spectral (Welch PSD band powers in delta, theta, alpha, sigma, and beta bands; delta-to-theta and theta-to-alpha ratios; spectral edge frequencies at 50% and 95%; spectral entropy), and two distributional (5th and 95th percentiles). These 140 signal features were concatenated with 12 algorithmic annotation features derived from CAISR outputs, including the automated apnea-hypopnea index, arousal index, limb movement index, sleep stage percentages, sleep efficiency, and model confidence scores, yielding a 152-dimensional feature vector. A gradient boosting classifier (300 estimators, max depth 5, learning rate 0.05) was trained with Standard-Scaler normalization. Demographic variables were deliberately excluded to reduce site-dependent confounding and improve cross-site generalization. On five-fold stratified cross-validation of the training set, our method achieved an accuracy of 0.773. On the unofficial validation set, we received a score of 0.632. Signal-derived spectral biomarkers, particularly EEG slowing ratios and sleep fragmentation indices, combined with automated sleep annotations, can predict cognitive impairment from PSG without demographic information, supporting a generalizable screening approach.