CausalEvidentialFuse: A unified framework for trustworthy multimodal clinical data fusion in healthcare diagnosis.

Authors

Abstract

Radiomics, the high-throughput extraction of quantitative imaging features from medical images, has transformed healthcare diagnosis by enabling the integration of heterogeneous clinical data sources such as electronic health records and medical imaging. However, existing multimodal fusion approaches suffer from critical trustworthiness limitations, including susceptibility to spurious cross-modal correlations, inadequate uncertainty quantification, and limited interpretability. These shortcomings impede clinical translation, where reliable and transparent predictions are essential. This study aims to develop a unified framework that simultaneously addresses causal robustness, uncertainty calibration, and clinical interpretability in multimodal clinical data fusion. To reflect the heterogeneous modality coverage of the four benchmark datasets used in this study-MIMIC-IV and eICU (structured EHR only), MIMICCXR (chest radiograph imaging plus radiomic features), and ADNI (MRI-derived radiomic features, cognitive scores, and APOE genotype)-we frame the framework as general multimodal clinical data fusion, with the radiomics-clinical fusion setting evaluated on the imaging-bearing cohorts (MIMICCXR and ADNI) and the genomic setting represented by the single APOE marker in ADNI. We propose CausalEvidentialFuse, comprising three synergistic components. First, a Cross-Modal Causal Attention module leverages structural causal models and do-calculus to disentangle genuine cross-modal relationships from confounding factors including scanner variability and demographic biases. Second, an Uncertainty-Aware Evidential Fusion module, built upon Dempster-Shafer evidence theory, quantifies both aleatoric and epistemic uncertainty using Dirichlet evidence distributions. Third, a Clinical Knowledge Graph-Guided Feedback (CKGF) mechanism incorporates domain knowledge from clinical knowledge graphs (SNOMED CT and ICD-10) to regularize predictions; this module does not involve direct clinician participation, prospective clinician feedback, or explicit clinical guideline supervision, and the previous term "Clinical Knowledge Graph-Guided Feedback (CKGF)" (CKGF) has therefore been revised throughout the manuscript to CKGF. The framework was evaluated on four large-scale, de-identified, publicly available benchmark datasets spanning diverse clinical domains: Medical Information Mart for Intensive Care (version IV), its Chest X-Ray collection, the Alzheimer's Disease Neuroimaging Initiative, and the electronic Intensive Care Unit Collaborative Research Database. CausalEvidentialFuse achieved the highest area under the receiver operating characteristic curve across all four datasets, with an average improvement of 2.5 percentage points (2.9% relative) over the best-performing baseline (EvidentialDL; mean AUC 0.899vs 0.874). The framework attained the lowest Expected Calibration Error of 0.031, indicating substantially better-calibrated uncertainty (a 35.4% relative reduction compared with EvidentialDL's ECE of 0.048, and a 54.4% reduction compared with the best non-evidential baseline MCAGN's ECE of 0.068). Epistemic uncertainty was reduced by 34.2% relative to baseline methods. Cross-dataset generalization demonstrated a performance retention rate of 91.8% (compared with 88.5% for the second-best method, EvidentialDL), and retrospective clinical validation confirmed clinically meaningful sensitivity and specificity across all diagnostic tasks. Inference time remained at 23 milliseconds, within clinical decision support requirements. CausalEvidentialFuse establishes a new standard for trustworthy multimodal healthcare diagnosis by integrating causal reasoning, evidential uncertainty quantification, and clinical knowledge guidance within a single unified framework. The results demonstrate that addressing multiple trustworthiness dimensions simultaneously yields complementary benefits-the causal module, evidential fusion module, and knowledge-guided feedback each contribute partially independent improvements that combine favorably-achieving superior diagnostic accuracy while providing the calibrated uncertainty estimates and interpretable reasoning chains required for confident clinical adoption.

Tags

Topics