Early prediction of cognitive decline enables timely intervention and improved clinical outcomes. In this work, our team, SleeperAgents, used overnight polysomnography (PSG) recordings to predict the onset of cognitive impairment, including mild cognitive impairment, Alzheimer's disease, and dementia, within 3–7 years of recording. We extracted 90 features across three categories: (1) demographic features (10 dimensions: age, BMI, encoded sex, and encoded race), (2) physiological features (49 dimensions) summarizing EEG, EOG, chin/leg EMG, ECG, respiratory, and SpO₂ signals via seven time-domain statistics (standard deviation, mean absolute value, zero-crossing rate, RMS, activity, Hjorth mobility, and complexity), and (3) algorithmic annotation features (31 dimensions) capturing respiratory event rates, sleep continuity metrics (TST, SOL, WASO, and REM latency), sleep stage distribution, transition entropy, limb movement rates, and annotation channel probability statistics. An XGBoost classifier was trained using leave-one-site-out cross-validation. To address the imbalance between the approximately 50% positive training distribution and the approximately 10% realistic test prevalence, out-of-fold predictions were used to fit an isotonic regression calibrator, producing well-calibrated probability estimates at inference. The model's performance, measured by AUROC, yielded values of 0.461 on the internal training set and 0.507 on the challenge validation set. Feature importance analysis revealed that the top-ranked XGBoost features were predominantly drawn from physiological and annotation-based categories, suggesting that PSG-derived sleep architecture and respiratory markers carry predictive signal for cognitive decline, though current discriminative performance remains limited. These initial results demonstrate the feasibility of a feature-based approach for cognitive decline prediction from PSG data and highlight the need for more advanced modeling approaches to enhance predictive accuracy. For the official phase, we plan to integrate deep learning architectures, such as sleep foundation models, with the existing feature-based pipeline to improve detection performance.