Objective: To determine whether combining four routine data sources improves in-hospital detection of lung cancer and heart failure, which sources matter, whether explanations are faithful, whether models are fair, and how much benefit federated training retains. Methods: This retrospective single-centre study linked MIMIC-IV admissions (73,637 patients) to chest radiographs, radiology reports, electrocardiograms (ECGs) and laboratory tests from the first 48 hours. Excluding prior cases, we compared single-source models, gradient boosting and early, late and attention-based fusion, tested explanations for faithfulness and weight dependence, audited subgroup performance, and simulated five hospitals for federated training. Results: Lung cancer was coded in 1.7% and heart failure in 15.6%. Attention fusion reached an area under the receiver operating characteristic curve (AUROC) of 0.833 (95% CI 0.809–0.859) for lung cancer and 0.885 (0.878–0.892) for heart failure, against 0.753 and 0.794 for gradient boosting on laboratory data (both p < 0.001); early and late fusion performed similarly. Lung cancer relied most on laboratory data and reports, heart failure on the ECG. Explanations passed a deletion test; weight randomisation of absolute attributions passed for ECG but not laboratory explanations. Heart-failure discrimination fell with age. Federated training recovered 95% of the centralised gain over local training for heart failure; for lung cancer, federated averaging (FedAvg) recovered 51% and FedProx 82%. Conclusion: Fusing routine data improved in-hospital detection of both diseases over laboratory data alone, and which data matter depends on the disease. Explanations should be checked per source, and performance in older patients monitored. External validation is needed.

Post-surgical adverse outcomes, including mortality, intensive care readmission, and complications, remain major challenges for clinical decision-making. Existing machine learning approaches focus on outcome prediction while operating as opaque systems, limiting clinical trust and the translation of predictions into treatment decisions, and many clinical studies rely on synthetic data in which shared intermediate variables create circular dependencies between inputs and targets that compromise reported performance. We aimed to develop an explainable multimodal architecture and a rigorous evaluation methodology that address these gaps. We designed a two-stage architecture integrating supervised deep learning for risk prediction with conservative Q-learning for action recommendation. The first stage uses five modality-specific encoders for structured records, physiological time-series, chest radiographs, clinical notes, and surgical metadata, unified through cross-modal attention into a shared patient-state representation. The second stage applies offline reinforcement learning to recommend clinical actions while preventing value overestimation. We formally characterized a target-leakage flaw in synthetic pipelines and propose a real-data methodology using a verified clinical database, with event-censored temporal separation and uncertainty-weighted per-task training. Component-level behavior was validated on a controlled synthetic benchmark, demonstrating that the architecture functions as designed without claiming clinical validity. The cross-modal attention and risk-prediction components behaved as expected, whereas the offline reinforcement learning stage did not converge on the benchmark, indicating that value estimation requires further investigation on real clinical data. The architecture provides dual-level explainability through attention visualization and value decomposition, contributing a deployable design, a formal methodological critique of synthetic-data practices, and a complete framework for clinically valid evaluation.