Interactive platform for supporting clinical decision-making using large language models (LLMs)
TL;DR
The results suggest that specialized, locally deployable fine-tuned models, combined with bounded interaction design, standardized output constraints, and workflow-oriented system integration, provide a secure, practical, and effective pathway for incorporating LLM-based decision support into routine pre-ambulatory clinical workflows while preserving safety, usability, and auditability in practice.
Abstract
Current healthcare systems face increasing workload, fragmented communication, and documentation burden, which contributes to delays and diagnostic errors in early triage. The proposed solution is intended to improve consistency, speed, and standardization in early patient assessments overall. This study presents an interactive clinical decision support platform that operationalizes a workflow-constrained, two-step history-taking process to support symptom-based differential diagnosis in the pre-ambulatory phase of care. The system addresses three tasks: (T1) automated patient history elicitation via constrained dialogue (exactly two multiple-choice follow-up questions), (T2) formalization of symptom narratives into structured medical (Latinate) terminology for clinician-facing documentation, and (T3) generation of differential diagnosis recommendations under a fixed and clinically interpretable output schema. We compare a generic large language model baseline (GPT-4, zero-shot) with a domain-adapted model (Llama-3 fine-tuned using LoRA) under identical interaction, prompting, and formatting constraints. Experiments on 903 symptom–diagnosis records and a held-out set of 200 controlled vignettes show that domain adaptation yields a 10–12% macro-F1 improvement and approximately 15% higher Recall on complex cases, while producing more clinically discriminative and diagnostically relevant follow-up questions as assessed by two medical raters . In a workflow-level evaluation, the platform reduced documentation time by 28% compared to standard intake and lowered the administrative effort required to prepare an initial clinician-facing summary for physician review. These results suggest that specialized, locally deployable fine-tuned models, combined with bounded interaction design, standardized output constraints, and workflow-oriented system integration, provide a secure, practical, and effective pathway for incorporating LLM-based decision support into routine pre-ambulatory clinical workflows while preserving safety, usability, and auditability in practice.