Aims: We set out to predict future cognitive impairment from polysomnography (PSG) recordings for the George B. Moody PhysioNet Challenge 2026. Most approaches to this task focus on EEG features, but EEG signal characteristics vary considerably across recording sites. We believe the ECG, which is consistent across sites, contains useful information about autonomic dysfunction in preclinical neurodegeneration that has been largely overlooked. Methods: Our baseline (team SleepLens) extracted 48 features from demographics and CAISR annotations (sleep stage proportions, arousal and respiratory indices, staging confidence, temporal shifts in N3 and REM) and trained an Optuna-tuned RandomForest. For the official phase, we are developing two sets of cardiac features. First, we use CAISR sleep staging to segment the ECG into stage-specific windows and compute HRV per stage: time-domain (RMSSD, SDNN), frequency-domain (LF/HF ratio), and nonlinear complexity (sample entropy, distribution entropy). The idea is that parasympathetic dysfunction during specific stages, especially NREM, may be an early marker of neurodegeneration that a single overnight HRV number would miss. Second, we treat CAISR-annotated arousals and respiratory events as natural cardiac stress tests: for each event, we extract heart rate acceleration at onset, peak change, and recovery time constant from a time-locked ECG window. A healthy autonomic system responds briskly. Sluggish or erratic recovery may reflect early network degradation. We also plan to explore pretrained EEG embeddings as a third input, combined via late fusion. Results: The baseline achieved AUROC 0.765 on cross-validation and 0.498 on the hidden validation set. The gap points to cross-site domain shift, which is part of why we are focusing on cardiac features that should generalize better. Conclusion: Stage-specific cardiac profiling and event-triggered autonomic responses may improve detection of cognitive decline from routine sleep studies, while avoiding the signal heterogeneity problems that limit EEG-based methods.