Fig. 1: Overview of the study.

Multiple cohorts are employed to evaluate our classification framework for multiple brain diseases including Alzheimer’s disease (AD), subcortical vascular mild cognitive impairment, and prediction of O[6]-methylguanine-DNA methyltransferase (MGMT) promoter methylation status. Our framework is evaluated for normal cognition (NC) vs. AD, static mild cognitive impairment (sMCI) vs. progressive MCI (pMCI), subcortical vascular mild cognitive impairment (svMCI) vs. subcortical vascular disease with no cognitive impairment (NCI), MGMT promoter methylation status (methylated or unmethylated) in aspects of classification performance in terms of area under the ROC curve (AUC), accuracy (ACC), sensitivity (SEN), specificity (SPE), and F1-score (F1) under 5-fold cross-validation, reliability evaluation, uncertainty analysis, ablation study, and generalizability evaluation. The framework is built on a two-stage training scheme: (1) Stage I (S1) aims to learn modality-specific disease-relevant feature representations using real multi-modal data, based on which the classification backbone and classification head of each modality are well established; (2) Stage II (S2) takes single-modal input and performs representation transfer to align the synthesized features with the reference ones from S1. To focus on feature transfer, the branch of available modality (marked in yellow) and the classification heads of all the modalities are borrowed from S1 and frozen in S2 (marked by gray background). To achieve efficient feature alignment, hierarchical similarity matching between the reference features (in S1) and synthesized features (in S2) in the classification backbone and classification head is imposed. More details are given in section “Two-stage training scheme and hierarchical feature similarity matching''.