Abstract
MedROAD V3 is a real-time FHIR R4 streaming platform that turns clinical sensor streams into calibrated deterioration risk, evaluated retrospectively on a cardiac ICU cohort. FHIR event sourcing, Kafka and Faust assemble an 86-element feature vector every five minutes. An XGBoost, LSTM and Transformer ensemble, combined by logistic meta-learning with SHAP attribution, scores each window; a write-back client returns the alert to the chart with a runtime-filtered language-model narrative. On 5049 held-out MIMIC-IV patients it reaches AUROC 0.816 at an Expected Calibration Error of 0.001, with a median inference path of 3.1 ms, so latency is governed by the EHR rather than by computation. The FHIR Observation resource is device-agnostic, so the pipeline ingests wearable streams unmodified. Three limits qualify these results. It is retrospective, and the remote-monitoring figures are a feature-restriction analysis on ICU data, not a wearable validation. Twenty-nine columns recording only how a patient was documented and managed, none physiological, reach AUROC 0.767 on the same split, so much of the discrimination reflects care process rather than physiology, although removing them costs only 0.007. Because OpenEMR 7.0.2 reference server grants no clinical FHIR write scope, the loop closes over an alternative path using its REST API, writing the alert as a Subjective-Objective-Assessment-Plan (SOAP) note on the encounter. The substitution works but is vendor-specific rather than portable, so the closed loop shown here is conditional on the target server.