Towards Explainable Doctor Recommendation with Large Language Models

Ziyang Zeng, Dongyuan Li, Yuqing YangarXiv 2025 · DOI · arXiv

Introduction

Online healthcare platforms often present patients with doctor recommendations based on simplistic metrics like consultation volume or star ratings. These "black box" systems frequently fail to capture the specific expertise required for complex medical conditions. Furthermore, the medical domain suffers from a "data desert," where expert-labeled data for training sophisticated ranking models is prohibitively expensive. To address these issues, researchers have developed a zero-shot, explainable framework that leverages Large Language Models (LLMs) to rank doctors based on nuanced medical needs without requiring massive labeled datasets.

Methodology: A Pointwise Ranking Framework

The proposed system utilizes a pointwise ranking framework, primarily powered by the Qwen2.5 series of LLMs. Unlike traditional listwise approaches that attempt to sort a group of candidates simultaneously, this framework evaluates each doctor individually against the patient's specific requirements.

Figure 1: Overview of the explainable doctor ranking framework, featuring dynamic criteria generation, forced elicitation for scoring, and multi-level explanations.

The framework introduces several key innovations:

  • Dynamic Ranking Criteria: The system prompts the LLM to generate "Evaluation Guidelines" tailored to the specific disease and treatment mentioned in the patient's query. This ensures that the ranking logic is medically relevant (e.g., prioritizing surgical success rates for lung cancer vs. longitudinal management for diabetes).
  • Forced Elicitation: To ensure consistency and extract numerical scores, the authors use specific prompt prefixes (e.g., "The professional relevance of the doctor is..."). This forces the model to output one of five predefined levels: Top, High, Mid, Low, or Not Relevant. Scores are then derived from the model's log-probabilities.
  • Two-Stage Explainability: Transparency is provided by displaying the generated ranking criteria and a detailed "Recommendation Explanation" for each doctor, fostering patient trust through interpretable reasoning.

The DrRank Dataset

To evaluate the framework, the authors introduced DrRank, a novel expertise-driven benchmark. It contains 38 unique disease-treatment pairs and 4,325 detailed doctor profiles. Ground truth annotations were provided by clinical experts from Peking Union Medical College Hospital to ensure high-quality relevance labels.

Figure 4: Distribution of professional relevance scores across various disease-treatment pairs in the DrRank dataset.

Beyond the specialized DrRank dataset, the framework was also tested on the BEIR benchmark (including TREC-COVID and Touche-2020) to demonstrate its generalization capabilities across broader scientific and argumentative retrieval tasks.

Key Insights and Results

The experimental results highlight the effectiveness of the LLM-based approach:

  • Performance: The framework using Qwen2.5-32B outperformed the strongest baseline by +6.45 NDCG@10 on the DrRank dataset, surpassing specialized cross-encoders pre-trained on massive general-domain data.
  • Logic over Format: Ablation studies revealed that the medical logic within the ranking criteria is the primary driver of performance. Replacing these with random but well-formatted criteria led to a significant drop in accuracy.
  • Capability Floor: While larger models (7B+) benefited significantly from the ranking criteria, smaller models (0.5B and 1.5B) struggled, suggesting a minimum capability threshold for handling such complex reasoning tasks.
  • Bias Reduction: By utilizing standardized criteria, the framework exhibited lower gender and geographic bias compared to traditional embedding-based models, which often pick up on irrelevant demographic cues in doctor biographies.

Limitations and Future Directions

Despite its success, the model faces certain challenges. A "prestige trap" was observed where the model occasionally over-ranked doctors from famous hospitals despite a lack of specific expertise in the target area. Future iterations may include "post-hoc calibration" using publication topics to mitigate this semantic drift. Additionally, the current system focuses on expertise but does not yet integrate practical constraints such as insurance compatibility, real-time availability, or system-wide workload balancing to prevent doctor burnout.

Summary updated

BibTeX
@misc{zeng2025towards,
  author = {Ziyang Zeng and Dongyuan Li and Yuqing Yang},
  title = {{Towards Explainable Doctor Recommendation with Large Language Models}},
  year = {2025},
  eprint = {2503.02298},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.02298},
}
Figure