A Reproducible Clinical Decision-Support Suite on MIMIC-IV
Most published clinical-AI results are single models on a single dataset, difficult to reproduce, and rarely validated outside their training hospital. We built a broad, methodologically rigorous, reproducible clinical decision-support (CDS) suite spanning four families - intensive-care deterioration and outcomes, emergency- department triage, electrocardiographic interpretation, and clini- cal natural-language processing - comprising 26 models. Tabular models are gradient-boosted trees over point-in-time, leakage- safe first-24-hour features; deep models include one-dimensional convolutional networks on raw 12-lead ECG, fine-tuned clinical transformers, and an instruction-tuned large language model for discharge-summary drafting. Every model uses patient-level data splits, probability calibration, a shuffled-label leakage gate, and SHAP explanations, and is characterised by its full confusion matrix with sensitivity, specificity and predictive values. Dis- crimination matched or approached published benchmarks: ICU mortality AUROC 0.884, acute kidney injury 0.830, prolonged stay 0.813; emergency-department-to-ICU 0.875; cardiologist- labelled ECG diagnosis 0.909; full-note diagnostic coding 0.892. Raw-signal ECG deep learning improved myocardial-infarction detection by +0.142 AUROC over interval features. The MIMIC- trained mortality model generalised to a different multi-centre US cohort (199,133 stays) with only a 0.044 AUROC drop. We describe how each model family is incorporated into the latest version of the zMed Critical Care application and its CDS tools