Multimodal Artificial Intelligence (AI) models-integrating diverse data such as imaging and clinical records-are advancing rapidly in healthcare, yet a significant disconnection persists between these complex predictive architectures and the explainable AI (XAI) techniques used to interpret them. We conducted a scoping review over 4 bibliographic databases to investigate the use of explainability methods in cross-modal medical AI studies. From 82 included studies, we found that the landscape remains dominated by independent feature attribution (assigning importance scores to individual modality in isolation), with the majority of studies relying on post-hoc methods (applied after a model decision is reached) that treat the model as a 'black box'. While emerging trends like visual grounding (linking textual justifications directly to specific image regions) and model reasoning show promise, a critical gap remains in explaining the underlying reasoning process. Standardised evaluation is missing in the majority of studies relying solely on qualitative measures. Only a minority of studies achieve good reproducibility with public codebase. We provide suggestions for the field to transition from individual and post-hoc XAIs toward intrinsically explainable designs where the reasoning logic is built directly into the model architecture to ensure that AI outputs align with human-centric clinical workflows and applications.
Kai Hu, Xing-Yue Fu, Yupeng Zhang et al.· npj Digital Medicine· 0 citations
Atrial fibrillation (AF), a common cardiac arrhythmia, presents significant challenges for early detection and management due to its asymptomatic and paroxysmal characteristics. In this study, we introduce the RetiAF score, a multimodal foundation-model-based biomarker derived from retinal fundus images for early detection of AF. We identified that the RetiAF score demonstrated a robust performance across multi-ethnic, multi-center datasets, achieving AUROCs of 0.8610 and 0.8019 on the UK Biobank (UKBB) development and internal testing datasets, and an AUROC of 0.7803 on an external dataset acquired in Shanghai, China. In addition, we identified that the RetiAF score consistently outperformed the traditional risk scores such as CHARGE-AF (AUROC: 0.7553) and C2HEST (AUROC: 0.7246) for the UKBB internal testing dataset. Multivariable logistic regression and propensity score analyses further demonstrated that the RetiAF score was independently associated with AF risk (p < 0.001). When stratified by higher C2HEST scores (≥3), the RetiAF score achieved an AUROC of 0.9619, highlighting its potential for identifying high-risk patients before the clinical onset of AF. The multimodal hybrid version of RetiAF (Hybrid_RetiAF) score, which incorporated clinical features (e.g., Age, BMI, etc) into the deep learning model, further enhanced predictive performance and achieved AUROCs of 0.8924 and 0.8381 on the UKBB Cohorts. As a secondary exploratory analysis, we evaluated whether RetiAF-derived scores were associated with chronic ischemic heart disease in UKBB, suggesting shared cardio-retinal risk information. These findings underscore the potentials of non-invasive retinal imaging as a scalable tool for AF and cardiovascular risk assessment, offering a promising alternative for large-scale screenings and personalized interventions.
Yupeng Xu, Yige Peng, Yan Jiang et al.· npj Digital Medicine· 0 citations