An explainable AI framework integrating FHIR R4 data normalization, physiological feature engineering, machine-learning risk stratification, and SHAP-based clinical explanation generation.
The Remote Patient Monitoring Clinical Decision Support (RPM-CDS) Framework is an explainable artificial intelligence framework designed to transform continuous patient-generated health data into structured clinical risk intelligence.
The framework integrates HL7 FHIR R4-based data normalization, physiological feature engineering, machine-learning risk stratification, explainable AI, and clinical decision-support workflows. It provides a standardized methodology for converting heterogeneous RPM data streams, including wearable sensors, home monitoring devices, and electronic health record observations, into actionable clinical insights.
The framework introduces an interpretable risk prediction pipeline that combines physiological feature modeling, machine-learning-based risk estimation, and SHAP-based explanation generation. Unlike conventional black-box prediction systems, the framework emphasizes explanation-as-a-clinical-artifact by representing model reasoning through interoperable clinical resources such as FHIR DiagnosticReport.
The core contribution of the RPM-CDS Framework is the integration of explainable machine learning with healthcare interoperability standards. The framework defines:
A structured dictionary containing 80 RPM-related features with LOINC references, UCUM units, and missing-data policies
Observation-based standardization of multi-stream RPM telemetry into interoperable FHIR resources
XGBoost-based risk prediction pipeline with probability calibration and risk score generation
Feature attribution and explanation generation using SHAP values mapped to clinical contexts
FHIR DiagnosticReport-based representation of model reasoning for clinical workflows
Clinical decision-support workflow integrating risk review, alerting, and clinical interpretation
Wearable devices · Home blood pressure monitoring · Continuous glucose monitoring · EHR laboratory observations
HL7 FHIR R4 · Observation Resource · LOINC terminology · UCUM units
RPM 80-Feature Dictionary · Cardiovascular · Respiratory · Metabolic · Activity · Patient-reported indicators · Data quality indicators
XGBoost risk model · Probability calibration · Risk score generation
SHAP feature attribution · Risk explanation · Feature contribution · Clinical interpretation
Clinical alerting · Risk review · Decision support workflow