Interpretable multimodal deep learning for diagnosis and symptom-onset prediction in genetic frontotemporal dementia

Background: Genetic frontotemporal dementia (FTD) is characterized by substantial clinical and neuroanatomical heterogeneity, making prediction of symptom onset challenging. We evaluated whether deep learning-derived MRI biomarkers could be combined with non-imaging biomarkers in an interpretable framework for diagnosis and phenoconversion prediction in genetic FTD.Methods: We evaluated Multi-ICNN, a multimodal extension of the Interpretable Convolutional Neural Network, using data from the FTD Risk Cohort. Imaging biomarkers derived from T1-weighted MRI were combined with demographic, genetic, and neuropsychological biomarkers in an additive prediction model. We evaluated classification of symptomatic mutation carriers versus genetically confirmed non-carriers and phenoconversion prediction within 2- and 5-year time horizons. Biomarker contributions were examined at the group and participant levels.Results: For diagnosis, Multi-ICNN achieved the highest mean area under the receiver operating characteristic curve (AUC) of 0.953, numerically outperforming the single-modality models. For phenoconversion prediction, the deep learning-based models outperformed the conventional volumetric and non-imaging models. The imaging-only ICNN achieved the highest mean AUCs of 0.798 and 0.780 for the 2- and 5-year horizons, respectively. Adding non-imaging biomarkers did not further improve ICNN performance. Neuropsychological measures contributed most strongly to diagnosis, whereas imaging-derived biomarkers dominated phenoconversion prediction, particularly biomarkers involving frontal and temporal regions. The additive model also provided participant-specific biomarker contributions.Conclusions: Interpretable deep learning-derived MRI biomarkers showed promise for phenoconversion prediction in genetic FTD, while the value of multimodal integration depended on the clinical task. The main limitations were the modest sample size and lack of external validation, reflecting the rarity of genetic FTD and the limited availability of comparable longitudinal cohorts.