To identify hippocampal radiomic patterns from structural MRI predicting Alzheimer’s biomarker status and compare the performance of radiomic, clinical, APOE4, and combined models.

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by cognitive decline and abnormal accumulation of beta-amyloid (Aβ) and tau proteins. Radiomics enables automated extraction of quantitative imaging features, capturing subtle textural patterns beyond visual assessment. This approach shows growing potential to improve diagnosis and prognosis in AD.

The study included 238 participants: 26 cognitively normal controls, 178 with mild cognitive impairment (MCI), and 34 with AD. SmallDependenceLowGrayLevelEmphasis and Skewness were the most important radiomic features across biomarkers. Combined models integrating radiomics, clinical data, and APOE4 achieved the best performance, with AUCs of 0.870 for Aβ42 and 0.844 for amyloid-PET. Removing APOE4 reduced the Aβ42 model’s AUC from 0.813 to 0.753. APOE4 had minimal impact on tau-PET or FDG-PET, where ADAS13 remained the strongest clinical predictor (AUC ≈ 0.77). Radiomic-only models also performed well for Aβ42 and tau-PET (AUC > 0.74). DALEX-based analysis identified SmallDependenceLowGrayLevelEmphasis, APOE4, and ADAS13 as the variables with the highest contribution to model performance, with mean dropout AUCs of 0.25, 0.22, and 0.30.