Authors - Gulshat Amirkhanova, Alikhan Amirkhanov Abstract - Wearable and Internet-of-Medical-Things (IoMT) devices now stream cardiac signals continuously, yet turning that flow into an early warning still needs a model that says when an adverse event is likely, not just whether one will occur. We treat that question as a time-toevent problem and benchmark five survival models on the UCI Heart Failure Clinical Records cohort (299 patients, 96 deaths, follow-up 4– 285 days): Cox proportional hazards, penalised Cox, random survival forest (RSF), gradient-boosted survival analysis, and the deep survival network DeepSurv. Models are compared under 5×5 repeated stratified cross-validation with Harrell’s and IPCW concordance, the integrated Brier score, and time-dependent AUC. The two tree-ensemble survival models lead: gradient boosting reaches a C-index of 0.731 and RSF 0.725, both ahead of Cox (0.709) and well ahead of DeepSurv (0.645), which also shows the worst calibration. On a cohort of this size the deep model does not pay its way. Permutation and SHAP analysis of the RSF point to serum creatinine, ejection fraction and age as the dominant risk drivers, which agrees with established cardiology. We frame the survival learner as the analytics stage of an IoMT cardiovascular-monitoring pipeline and release code, data split and seeds for full reproducibility.
