JMIR Publications

ABSTRACT

Background:

Machine-learning systems can identify patients at increased risk of hospital readmission, but predictive performance alone does not ensure that model outputs can be appropriately interpreted or translated into safe clinical action. Effective clinical decision support requires integration of predictive evidence with explanation, contextual reasoning, recommendation generation, and appropriate professional oversight.

Objective:

This study aimed to develop and demonstrate an integrated architecture that connects predictive analytics, explainability, semantic reasoning, recommendation generation, and clinician oversight within a unified clinical decision-support workflow.

Methods:

A design-oriented framework was implemented using a synthetic demonstration dataset and integrated machine-learning prediction with SHAP and LIME explanations, the established Medication Safety Co-Decision Ontology (MSCDO), ontology-based reasoning, recommendation generation, and clinician review. The architecture additionally incorporated uncertainty, fairness, traceability, data governance, and human oversight as cross-cutting trustworthiness considerations. The resulting workflow was demonstrated through a patient-specific scenario in which computational outputs were progressively transformed into semantically enriched information, reasoning states, potential actions, and clinician-controlled decisions.

Results:

The integrated architecture successfully connected predictive outputs with global and patient-specific explanations, semantic representations, ontology-based reasoning, and structured recommendations. In the illustrative patient scenario, an 87% predicted probability was classified as high risk, accompanied by a displayed prediction confidence of 92%. The semantic layer incorporated major contributing factors and linked them to reasoning states involving medication-related risk, polypharmacy, clinical vulnerability, and enhanced monitoring. The framework generated potential actions including medication review, pharmacist consultation, enhanced follow-up, care coordination, and discharge planning. Clinicians had explicit options to accept, modify, or override the generated outputs.

Conclusions:

The study demonstrates the technical feasibility of integrating predictive analytics, explainability, semantic reasoning, and professional oversight within a single trustworthy clinical decision-support architecture. Rather than positioning artificial intelligence as an autonomous decision-maker, the framework establishes a structured pathway for clinicians to interpret, contextualize, and review computational evidence. External validation, quantitative assessment of fairness and uncertainty, clinician-centered evaluation, and prospective assessment of clinical utility are required before establishing effectiveness or patient-level benefit.