JMIR Publications

ABSTRACT

Background:

Clinical decision-support systems usually assume electronic health records rich in vital signs and coded diagnoses. Hospital-at-home (HaH) programmes often rely instead on a hand-maintained daily census, where evidence of deterioration, prolonged stays, and missing safety decisions is scattered across days and free text. We frame daily census review as longitudinal, capacity-constrained triage and propose an ontology-guided multi-agent framework in which four cooperating agents turn unstructured operational records into an auditable, ranked worklist. An intake agent reconstructs care episodes and maps free-text diagnoses to the Human Disease Ontology; specialist Stay, Risk, and Safety agents combine supervised learning, large-language-model (LLM) reading, and explicit clinical rules; a coordinator agent ranks patients within review capacity. We validate every architectural component through a five-part ablation study on 23 months of Johns Hopkins Aramco Healthcare data (N=,1,404 episodes, 20,290 patient-days). Removing episode-trajectory features reduces Stay-agent AUC by 0.070 and Risk-agent AUC by 0.064 (five-fold CV), confirming that longitudinal signals are necessary. Replacing the programme's static admission-day risk score with dynamic daily reassessment improves AUC from 0.514 to 0.721 (+0.207 at 1,705 decision points). The Safety Agent's rule-judge architecture achieves precision=1.00 at 0.2 alerts per day, versus precision=,0.11 and 0.6 alerts per day when an LLM judges documentation gaps directly an 89% false-positive reduction. The multi-agent coordinator recovers 78.9% of documentation-gap episodes within a daily review budget of 30, compared with 5.3% for a monolithic any-problem model a 15-fold improvement while maintaining full precision (all 30 flagged episodes carry at least one confirmed adverse event). Stay-agent held-out AUC is 0.849; Risk-agent held-out AUC is 0.632. These results demonstrate that ontology-grounded, ablation-validated multi-agent coordination can transform sparse, text-heavy HaH records into safe, auditable clinical decision support without requiring structured EHRs. The system is intended to support, rather than replace, clinical judgment. Code and results: \url{https://github.com/SarahayadAOU/HomeCare_JHA.

Objective:

This study aimed to develop an ontology-guided multi-agent framework that frames daily census review as longitudinal, capacity-constrained triage and turns unstructured operational records into an auditable, ranked worklist, and to evaluate the contribution of each architectural component.

Methods:

In the framework, an intake agent reconstructs care episodes and maps free-text diagnoses to the Human Disease Ontology; specialist Stay, Risk, and Safety agents combine supervised learning, large language model (LLM) reading, and explicit clinical rules; and a coordinator agent ranks patients within review capacity. We validated every architectural component through a five-part ablation study on 23 months of retrospective daily census data from the Johns Hopkins Aramco Healthcare HaH programme in Saudi Arabia (1,627 episodes, of which 1,404 had a known outcome; 20,290 patient-days). Discrimination was reported as the area under the receiver operating characteristic curve (AUC), and system-level performance as recall and precision within a fixed daily review budget.

Results:

Removing episode-trajectory features reduced Stay-agent AUC by 0.070 and Risk-agent AUC by 0.064 (five-fold cross-validation), confirming that longitudinal signals are necessary. Replacing the programme’s static admission-day risk score with dynamic daily reassessment improved AUC from 0.514 to 0.721 (+0.207 at 1,705 decision points). The Safety Agent’s rule-judge architecture achieved precision = 1.00 at 0.2 alerts per day, versus precision = 0.11 and 0.6 alerts per day for a simulated LLM-alone baseline that judges documentation gaps directly, an 89% false-positive reduction. The multi-agent coordinator recovered 78.9% of documentation-gap episodes within a daily review budget of 30, compared with 5.3% for a monolithic any-problem model, a 15-fold improvement, while maintaining full precision (all 30 flagged episodes carry at least one confirmed adverse event). Stay-agent held-out AUC was 0.849; Risk-agent held-out AUC was 0.632.

Conclusions:

These results demonstrate that ontology-grounded, ablation-validated multi-agent coordination can transform sparse, text-heavy HaH records into safe, auditable clinical decision support without requiring structured electronic health records. The system is intended to support, rather than replace, clinical judgment.