C-PATH: Conversational Patient Assistance and Triage in Healthcare System

Qi Shi, Qiwei Han, Cláudia Soares

Abstract

Navigating healthcare systems can be complex and overwhelming, creating barriers for patients seeking timely and appropriate medical attention. In this paper, we introduce C-PATH (Conversational Patient Assistance and Triage in Healthcare), a novel conversational AI system powered by large language models (LLMs) designed to assist patients in recognizing symptoms and recommending appropriate medical departments through natural, multi-turn dialogues. C-PATH is fine-tuned on medical knowledge, dialogue data, and clinical summaries using a multi-stage pipeline built on the LLaMA3 architecture. A core contribution of this work is a GPT-based data augmentation framework that transforms structured clinical knowledge from DDXPlus into lay-person-friendly conversations, allowing alignment with patient communication norms. We also implement a scalable conversation history management strategy to ensure long-range coherence. Evaluation with GPTScore demonstrates strong performance across dimensions such as clarity, informativeness, and recommendation accuracy. Quantitative benchmarks show that C-PATH achieves superior performance in GPT-rewritten conversational datasets, significantly outperforming domain-specific baselines. C-PATH represents a step forward in the development of user-centric, accessible, and accurate AI tools for digital health assistance and triage.

Introduction

Modern healthcare navigation presents a significant challenge for the average patient. Traditional digital symptom checkers often rely on rigid, rule-based logic or utilize complex medical terminology that exceeds typical health literacy levels. While Large Language Models (LLMs) offer a more intuitive, natural language interface, they are frequently prone to "hallucinations" and often lack alignment with how patients actually describe their ailments in a clinical setting.

C-PATH (Conversational Patient Assistance and Triage in Healthcare) addresses these gaps by providing a medically grounded yet conversationally fluent interface. The system guides users from initial symptom reporting to specific department recommendations and specialist referrals, ensuring that the transition from self-assessment to professional care is seamless and accurate.

Figure 1: Overview of the C-PATH workflow, illustrating the transition from symptom checking to department recommendation and specialist referral.

Methodology

Architecture and Training Pipeline

The C-PATH system is built upon the LLaMA3-8B architecture, utilizing a specialized three-stage fine-tuning pipeline designed to balance deep medical knowledge with conversational accessibility.

  1. Knowledge Injection: The model is first exposed to large-scale medical QA datasets, including PubmedQA and MedQA-USMLE, to establish a foundational understanding of clinical concepts.
  2. Instruction Tuning: To transition from static question-answering to dynamic interaction, the model is trained on doctor-patient dialogue datasets.
  3. Summarization Tuning: The final stage focuses on generating concise, EHR-compatible clinical summaries based on the dialogue, facilitating efficient hand-offs to human medical professionals.
Figure 2: The C-PATH multi-stage fine-tuning pipeline, integrating medical knowledge, dialogue data, and clinical notes using the LMFlow framework.

Technical Innovations

A primary challenge in medical AI is the gap between structured clinical data and patient descriptions. C-PATH utilizes a GPT-based data augmentation framework to bridge this gap. By taking structured data from the DDXPlus dataset (comprising 1.3 million cases), the team used GPT-3.5 Turbo to rewrite clinical jargon into natural language. For example, technical terms like "iliac wing(R)" were translated to "right side of my lower back," ensuring the model recognizes symptoms as they are naturally reported.

To maintain coherence in long, multi-turn dialogues, C-PATH employs Scalable History Management. This includes Context Window Pruning (a sliding window of recent turns) and Turn-Level Summarization, which condenses older parts of the conversation into a brief system summary to stay within token limits without losing vital context.

Results and Discussion

Triage and Recommendation Accuracy

C-PATH was benchmarked against several baselines, including domain-specific models like ClinicalBERT. On GPT-rewritten datasets, C-PATH achieved an exceptional F1-score of 0.996 for department recommendations. This suggests that the model's conversational fluency does not come at the cost of diagnostic accuracy.

GPTScore Evaluation

Evaluation using GPTScore highlighted the model's strengths in Informativeness (100.00) and Accuracy (100.00). While the "layperson translation" resulted in a slight decrease in technical specificity, it significantly improved Understandability (94.57 vs. 76.26 for raw data). This trade-off is essential for patient-facing tools, where clarity is a prerequisite for effective triage.

Key Insights

The study demonstrates that a fine-tuned generalist model (LLaMA3) can outperform specialized clinical models in conversational tasks. The normalization of departmental variations through GPT-rewriting proved vital for stable training, suggesting that data quality and linguistic alignment are as important as model size in the medical domain.

Conclusion and Future Work

C-PATH represents a significant advancement in accessible digital health assistance. However, limitations remain, including susceptibility to logical pitfalls common in LLMs (such as the "Reversal Curse") and dataset imbalances favoring respiratory and general medicine.

Future iterations will focus on incorporating Reinforcement Learning from Human Feedback (RLHF) with clinical expert supervision. This will further align the model with professional medical standards, reduce hallucination risks, and expand the system's efficacy across a broader range of medical specialties.

Summary updated

Abstract

Navigating healthcare systems can be complex and overwhelming, creating barriers for patients seeking timely and appropriate medical attention. In this paper, we introduce C-PATH (Conversational Patient Assistance and Triage in Healthcare), a novel conversational AI system powered by large language models (LLMs) designed to assist patients in recognizing symptoms and recommending appropriate medical departments through natural, multi-turn dialogues. C-PATH is fine-tuned on medical knowledge, dialogue data, and clinical summaries using a multi-stage pipeline built on the LLaMA3 architecture. A core contribution of this work is a GPT-based data augmentation framework that transforms structured clinical knowledge from DDXPlus into lay-person-friendly conversations, allowing alignment with patient communication norms. We also implement a scalable conversation history management strategy to ensure long-range coherence. Evaluation with GPTScore demonstrates strong performance across dimensions such as clarity, informativeness, and recommendation accuracy. Quantitative benchmarks show that C-PATH achieves superior performance in GPT-rewritten conversational datasets, significantly outperforming domain-specific baselines. C-PATH represents a step forward in the development of user-centric, accessible, and accurate AI tools for digital health assistance and triage.
BibTeX
@inproceedings{shi2025c,
  author = {Qi Shi and Qiwei Han and Cláudia Soares},
  title = {{C-PATH: Conversational Patient Assistance and Triage in Healthcare System}},
  booktitle = {ICDH},
  year = {2025},
  doi = {10.1109/icdh67620.2025.00016},
}
Figure