For the PhysioNet Challenge 2026, Leicester Fox team developed a mul-timodal late-fusion pipeline for patient-level prediction from polysomnogra-phy (PSG) and structured clinical metadata, designed to predict future diag-noses of cognitive impairment. The model separates metadata and signal processing into two independent-ly trained branches and fuses them at the probability level, enabling branch-specific preprocessing, augmentation, and interpretability. A gradient boost-ing classifier is trained on this representation with synthetic metadata mask-ing and occasional full-row dropout to improve robustness to incomplete inputs. The signal branch operates on physiological EDF recordings and CAISR algorithmic annotation EDFs. Channels are standardized by alias matching and bipolar derivation where needed. Seven lead groups are repre-sented: EEG, EOG, chin EMG, leg EMG, ECG, respiration, and SpO2. For each available group, compact statistical and dynamical descriptors are ex-tracted, including amplitude, variance, zero-crossing, and Hjorth features. These are combined with CAISR-derived event-density and sleep-architecture summaries. Signal extraction is time-structured, using two fixed windows (1200 s and 2400 s from 1800 s) and three anchor windows at 20%, 50%, and 80% of recording duration, plus a full-night CAISR summary block. The resulting signal vector is modeled with extremely randomized trees.Signal training uses structured augmentation with Gaussian noise, multiplicative jitter, feature dropout, and block dropout across windowed and full-night feature groups. Final prediction is obtained as 0.60 * p_meta + 0.40 * p_signal. Local evaluation with 5-fold stratified cross-validation yielded average AUROC 0.8445, accuracy 0.7231, and F1 0.6805. External Challenge vali-dation yielded AUROC 0.609, AUPRC 0.161, accuracy 0.671, and F-measure 0.090, ranked at 54/244. Overall, this multimodal late-fusion framework demonstrates strong internal performance and robustness to missing data but shows limited generalization on external data.