An Exploratory Six-Probe Blood RNA Signature for Predicting 12-Month Cognitive Decline Along the Alzheimer's Disease Continuum: An Interpretable Machine Learning Study

Background/Objectives: Predicting how fast a patient with Alzheimer's disease will decline over the next year remains a challenge. Existing blood transcriptomic studies have not established whether probe selection is reproducible, whether the signal is transcriptional or reflects immune cell shifts, or whether they generalise across platforms. Methods: We applied five steps to 96 ADNI-GO whole-blood microarray samples (Affymetrix HG-U219; 12-month MMSE change): PyImpetus Markov Blanket selection, Elastic Net with leave-one-out cross-validation (LOOCV), SHAP attribution, MCP-counter cell-type deconvolution, and cross-platform mapping into AddNeuroMed (GSE63060, n = 329, Illumina). Feature selection preceded cross-validation without constituting data leakage. Results: The same six probes emerged across four independent runs (Jaccard J = 0.214, p = 0.03): AQP7, RPS5, CHD2, SNX5, ASS1, and an uncharacterised chr12q15 transcript. The panel achieved LOOCV MAE = 1.388 and R² = 0.247, outperforming the full-probe baseline by 14.9%. All probes survived immune cell correction with signs intact. SNX5 replicated in AddNeuroMed (r = -0.170, p = 0.002). Conclusions: The exploratory six-probe blood RNA panel predicts 12-month cognitive decline (LOOCV R² = 0.247) with transcriptional origin confirmed by cell-type deconvolution and cross-platform evidence for SNX5. External testing in ADNI-2 (n = 91, R² = -0.222) showed that generalisation depends on visit-timepoint matching, indicating clinical utility cannot yet be claimed and defining conditions for prospective validation. Code and a research prototype tool are publicly available.