Olfactory radiomics signatures link pTau217 to cognitive impairment in probable Alzheimer’s disease: a multi-cohort machine learning study

Alzheimer's disease (AD) diagnosis remains challenging because current molecular biomarkers, though sensitive, lack spatial specificity. Radiomics offers a promising alternative by quantifying subtle microstructural alterations from routine MRI that are invisible to traditional volumetric analyses. Given that olfactory-related regions are preferentially affected in AD with olfactory dysfunction preceding cognitive decline, this study aimed to develop and validate radiomics signatures from these regions as imaging biomarkers linking molecular pathology to clinical phenotypes. Radiomics models were developed and validated across 834 participants from three independent cohorts: in-house cohort of 278 participants (122 AD, 156 cognitively unimpaired controls) ADNI and OASIS validation (each with n = 278). A total of 1,502 radiomics features were extracted from each bilateral olfactory-related brain region using structural MRI, with optimal features selected through correlation analysis, recursive feature elimination, and LASSO regression. Six machine learning algorithms were systematically trained and validated, with associations between radiomics signatures, plasma pTau217, olfactory identification, and cognitive performance examined through partial correlation and mediation analyses. Direct comparisons with volumetric models and hierarchical regression analyses were performed to quantify the added value of radiomics. The hippocampus-amygdala radiomics signatures achieved superior diagnostic performance across all six algorithms and three cohorts (AUC: 0.86–0.92, accuracy: 0.81–0.88). These signatures significantly outperformed those from piriform cortex, entorhinal cortex, and orbitofrontal cortex. Radiomics models demonstrated marginally higher AUC values and superior sensitivity (0.776–0.869 vs 0.505–0.811) compared to volumetric models. Hierarchical regression analyses revealed that radiomics features showed significant incremental contributions in 75.7% of analyses compared to only 12.5% for volumetric features in the all-participants group (6.1-fold advantage), with substantially higher mean incremental variance explained (ΔAdjR2: 0.022 vs 0.004, p < 0.001); across all combined analyses, radiomics features maintained a 2.2-fold advantage (31.4% vs 14.2%). In the in-house cohort (n = 278), where standardized plasma biomarker and olfactory assessments were available, the radiomics score showed negative correlations with global cognition (r = -0.28, p < 0.05) in AD patients, and mediation analyses further revealed that hippocampal-amygdala radiomics features statistically mediated 26.9%-37.1% of the tau-cognition association and 26.6%-44.2% of the olfaction-cognition association. Radiomics signatures from olfactory-related regions, particularly the hippocampus-amygdala complex, represent non-invasive associative imaging biomarkers statistically linking tau pathology and olfactory deficits to cognitive impairment. These findings establish accessible imaging correlates that complement blood tests for enhanced diagnosis and patient stratification in AD.