Abstract
INTRODUCTION: Machine learning applied to neuroimaging can help with medical diagnosis and early detection by identifying biomarkers of subtle changes in brain structure and function. The effectiveness of advanced diffusion MRI (dMRI) methods for pre-dementia classification remains largely unexplored, particularly when combined with CSF biomarkers. METHODS: We implemented XGBoost machine learning models to evaluate the classification potential of dMRI parameters (derived using NODDI, C-NODDI, MAP, or SMI), CSF biomarkers of Alzheimer's pathology (Tau, pTau, Aβ42, Aβ40), and pairwise dMRI + CSF combinations in distinguishing cognitive normality from mild cognitive impairment. RESULTS: MAP-RTAP (AUC = 0.78) and pTau/Aβ42 (AUC = 0.76) were the best performing individual biomarkers. Combining C-NDI derived using C-NODDI and Aβ42/Aβ40 achieved the highest performance (AUC = 0.84) and accuracy (0.84), while other combinations optimized either sensitivity (0.93) or specificity (0.88). DISCUSSION: dMRI biomarkers demonstrate comparable performance to CSF biomarkers, with notable improvements achieved when combined. This study highlights dMRI's effectiveness for enhancing early AD detection. HIGHLIGHTS: Advanced multishell diffusion MRI provides equivalent performance as CSF biomarkers in classifying MCICombining diffusion MRI and CSF biomarkers improves classification performanceStatistical diffusion MRI models perform best when used individually to classify MCIThe pTau/Aβ42 ratio outperforms other individual CSF biomarkers in MCI diagnosisBiophysical diffusion MRI models achieve the best performance when combined with CSF data.