Data-centric, robust, and explainable multimodal deep learning for clinical decision support: A systematic review.
TL;DR
Within this recent, self-selected slice of the field, technical robustness and explainability techniques are widely reported but rarely validated out-of-distribution or against clinical outcomes, and the underlying evidence is at high risk of bias.
Abstract
Purpose
Multimodal deep learning is increasingly proposed for clinical decision support (CDS) under a "data-centric" framing that prioritizes label quality, missing-modality robustness, distribution shift, calibration, and explainability. Prior reviews have examined multimodal medical AI, CDS, and data-centric methods separately, but none address their intersection. We mapped the modalities, fusion strategies, and data-centric and explainability techniques used in this recent literature, quantified how often each is implemented rather than merely mentioned, assessed deployment-relevant evidence (external validation, clinical-outcome measurement, equity), and formally appraised study-level risk of bias.
Methods
Following the PRISMA 2020 statement (PROSPERO CRD420261427815; registered retrospectively), we screened 150 records and included primary, clinical, multimodal studies that applied machine or deep learning to a decision-support task and reported at least one quantitative result. Two reviewers screened and extracted data with consensus adjudication. Each study was coded against pre-specified operational definitions, separating implemented or empirically evaluated techniques from those only mentioned. Study-level risk of bias was assessed with PROBAST + AI. Synthesis was narrative.
Results
Thirty-one studies met inclusion; 30 (97%) were published between 2024 and 2026, with a median of three modalities (range 2-6), most commonly structured EHR (71%) and imaging (39%). Data-centric techniques were frequently reported (74-84% across label-noise, distribution-shift, calibration, missing-modality and class-imbalance handling; equity 61%). However, external validation was reported in only 4/31 studies (13%), a clinical or provider outcome in 3/31 (10%), and no study reported routine deployment. Overall risk of bias was high in 27/31 studies (87%), driven by the analysis domain.
Conclusion
Within this recent, self-selected slice of the field, technical robustness and explainability techniques are widely reported but rarely validated out-of-distribution or against clinical outcomes, and the underlying evidence is at high risk of bias. Progress requires external multi-site validation, clinical-outcome measurement, formal bias appraisal, and adherence to AI reporting standards (e.g., TRIPOD + AI) before deployment can be justified.