TRIAGE: Dialectical Reasoning for Explainable Risk Prediction on Irregularly Sampled Medical Time Series with LLMs

Researchers introduce TRIAGE, an LLM-based framework that uses dialectical reasoning to improve risk prediction on irregularly sampled medical time series data. The approach generates competing clinical outcome rationales to produce calibrated, continuous risk scores rather than overconfident binary predictions, achieving 3.3% AUPRC improvement and 81% reduction in calibration error versus baseline methods.

TRIAGE addresses a critical gap in clinical decision support systems where large language models tend to collapse nuanced risk assessments into binary predictions, undermining both calibration accuracy and cross-patient comparability. This research is significant because medical triage fundamentally depends on granular risk stratification—clinicians need to distinguish not just between sick and well patients, but across multiple severity levels. The dialectical reasoning approach, which elicits competing outcome-specific rationales, forces the model to explore clinical evidence supporting different scenarios rather than committing prematurely to a single outcome prediction.

The technical innovation builds on growing interest in making LLM decision-making more interpretable and clinically defensible. Prior work explored LLMs for medical time series analysis but struggled with overconfidence and poor calibration. TRIAGE's framework generates continuous risk scores anchored in explicit clinical reasoning, enabling clinicians to understand and verify the model's logic. The 81% calibration error reduction is particularly important, as miscalibrated systems mislead clinicians about actual risk levels, potentially causing diagnostic errors.

This work impacts clinical AI development by demonstrating that dialectical reasoning improves both performance metrics and human interpretability. The results suggest that forcing LLMs to reason through competing hypotheses produces more robust, verifiable outputs than direct prediction approaches. For healthcare providers and AI developers, this framework could enhance adoption of LLM-based clinical tools by addressing longstanding concerns about black-box decision-making and overconfident predictions. Future developments likely involve validating TRIAGE across additional medical domains and integrating it into real-world EHR systems where clinician feedback can further refine reasoning quality.

• →TRIAGE uses dialectical reasoning to generate competing clinical outcome rationales, mitigating risk polarization in LLM predictions.
• →The framework achieves 3.3% average AUPRC improvement and reduces calibration error by 81% compared to competitive baselines.
• →Continuous risk scores grounded in explicit clinical reasoning improve interpretability and clinician verification compared to binary predictions.
• →LLM-as-a-judge evaluation shows TRIAGE's rationales exceed post-hoc explanation quality by 20% in clinical reasoning assessment.
• →The approach addresses critical limitations in using LLMs for medical time series analysis on irregularly sampled electronic health record data.