Alzheimer's Disease

AD-related GeoSAE features contributed most to MCI-to-AD conversion prediction. After additional modifications, the Alzheimer’s scenario reproduced reported patterns of ceramide accumulation, sphingosine accumulation after an initial drop, and decreased S1P species. StackFeat-RL achieved the highest mean AUC across all three Alzheimer's disease tasks while using smaller panels than ElasticNet and Boruta. The Definite AD panel was linked to mechanisms including autophagy, oxidative stress, neuroinflammation, BDNF signalling, hypoxia, and haemoglobin or haeme pathways. In the strict Alzheimer’s subset, tortuosity_4x increased mean AD probability relative to baseline. Comorbidity-only features did not meaningfully predict MCI-to-AD conversion. For Normal vs Probable AD, StackFeat-RL significantly outperformed ElasticNet while using far fewer genes. In the Alzheimer’s cohort, Bézier features outperformed BiomedCLIP and image-level EfficientNet-B2 on test AUC. The results support a separation between AD-related and comorbidity-related signal. Alzheimer’s disease cannot be fully explained by analyzing a single factor alone. Alzheimer’s Spearman analysis was power-limited, but std_curvat…