MRI-based radiomics and volumetrics for predicting Alzheimer's disease onset with explainable machine learning

• MRI-based radiomics and volumetrics combined with explainable machine learning can predict the onset of Alzheimer’s disease.¹
• The study is published in NeuroImage Volume 329 (April 2026), by Louise Bloch, Katarzyna Borys, Felix Nensa, Christoph M. Friedrich, and the Alzheimer’s Disease Neuroimaging Initiative.¹
• The approach is structural/neuroimaging-only and is tangential to electrophysiology and BCI; tier-2 for neuroimaging methods.^{1 1}

Weekly enrichment (2026-07-20)

• The paper (NeuroImage, 2026; DOI 10.1016/j.neuroimage.2026.121813) frames Alzheimer’s prediction as a time-to-event (survival) problem, forecasting onset for cognitively normal (CN) and mild-cognitive-impairment (MCI) subjects rather than classifying disease stage at scan time, and handles right-censored follow-up.²
• Data pipeline: T1-weighted MRI were segmented into 95 brain regions with FastSurfer v1.1.2 (Desikan-Killiany-Tourville atlas), then PyRadiomics v3.0.1 extracted 107 features per region (first-order, GLCM, GLRLM, GLSZM, NGTDM, GLDM, and shape), yielding 10,165 radiomics features complemented by 41 clinical baseline features.²
• Cohorts: models were trained and internally validated on ADNI (1,622 subjects after selection), with external validation on AIBL (322 subjects) and OASIS-3 (445 subjects); features were field-strength-corrected via residual adjustment to handle 1.5 T vs. 3 T scans.²
• Five survival models were compared — Cox proportional hazards, Coxnet/elastic-net Cox, Extra Survival Trees, Gradient Boosting Survival, and Random Survival Forest (scikit-survival), with sample weighting to offset low conversion rates.²
• Performance: radiomics-based feature sets improved eight-year prediction Brier scores by roughly 0.11 to 3.02 percentage points over comparison models, with the best integrated Brier score around 8.12% (Gradient Boosting Survival, baseline + radiomics) versus about 8.85% for baseline-only features.²
• Explainability used SHAP plus “high-level” interpretations, and regional relevance was checked against Voxel-Based Morphometry (VBM) to argue biological plausibility.²
• Key radiomic signatures linked to higher AD risk included complex texture in the left entorhinal cortex, an irregular shape of the right amygdala, and fine-granular texture of the left middle temporal gyrus.²
• Scope note: this is a structural-neuroimaging and machine-learning study with no electrophysiology or brain–computer-interface component; it remains tier-2 for BCI relevance, and reproducible code is released by the authors.²³

Footnotes

1. https://www.sciencedirect.com/science/article/pii/S105381192600131X?dgcid=rss_sd_all ↩ ↩² ↩³ ↩⁴

2. https://doi.org/10.1016/j.neuroimage.2026.121813 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://github.com/LouiseBloch/AlzheimersDiseaseExplainableMRIRadiomics ↩