We present a clinically grounded pipeline for predicting future cognitive impairment from polysomnographic (PSG) recordings, developed for the George B. Moody PhysioNet Challenge 2026. In the first stage, we systematically review clinical evidence linking cognitive impairment and its precursors including disordered sleep breathing, autonomic dysregulation, and early neurodegenerative markers to measurable physiological changes during sleep. In the second stage, we identify the specific PSG signatures through which these conditions manifest, constructing a physiologically motivated feature set that connects observable sleep biomarkers to long-term cognitive outcomes. Beyond classical summary statistics derived from sleep architecture and respiratory events, we investigate Markov Transition Fields (MTFs) as a complementary representation. MTFs encode the sequential dynamics of sleep stage transitions and physiological signal evolution as two-dimensional image matrices, capturing temporal dependencies and recurrence patterns that may reflect subtle alterations in sleep architecture associated with incipient cognitive decline. We further explore self supervised pre training on raw PSG signals to derive subject level sleep fingerprints, which are compact latent representations encoding individual physiological patterns without requiring labelled data. These representations may capture idiosyncratic deviations in sleep microstructure that serve as early markers of neurodegeneration. In parallel, we investigate transformer based architectures operating directly on multi-channel PSG sequences, leveraging attention mechanisms to model long range dependencies across the full overnight recording. We compare classical machine learning models on the engineered feature set against self supervised and transformer based approaches, applying site aware cross-validation throughout. Results will be reported upon completion of model training and validation