PRISMATIC: Precision Risk Inspection System for Multi-Agent Tactical Interaction in Clinical Decision

ACL ARR 2025 May Submission6883 Authors

Abstract:

Medication prescribing errors remain a critical challenge in clinical practice, often stemming from incomplete patient understanding, ambiguous documentation, and suboptimal decision support.
In this paper, we propose PRISMATIC, a 3-layer multi-agent prescription risk mitigation framework designed to generate safe, interpretable, and traceable drug regimens by analyzing unstructured patient clinical note texts. Our implementation of PRISMATIC is available at https://anonymous.4open.science/r/PRISMATIC.
To enhance adaptability and safety, PRISMATIC integrates two mechanisms: (1) Dynamic Self-updating Weight Adjustment (DSWA), which tunes risk factor weights over time, and (2) Differential Feedback Calibration Mechanism (DFCM), which learns from discrepancies with gold-standard prescriptions to improve future outputs.
Evaluated on a curated dataset from MIMIC-IV, PRISMATIC outperforms raw LLM outputs and prompting-based baselines (Few-Shot, Chain-of-Thought, ReAct, Tree-of-Thoughts) in reducing prescription risks. These results highlight the potential of multi-agent systems for improving clinical medication decision support.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: healthcare applications, clinical NLP, LLM/AI agents;

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 6883