TL;DR
This paper introduces a novel deep learning framework, Hybrid-GRUCNN-Attention, for matching patients with the most suitable medical experts in online Q&A communities. By integrating diverse patient and physician data with advanced neural network architectures, the model significantly improves recommendation accuracy and efficiency, addressing a critical need for better resource allocation in digital healthcare.
What Problem Does This Paper Address?
Online healthcare Q&A communities are invaluable for patient education and physician-patient communication, but they face significant challenges. The sheer volume of users and queries leads to a mismatch between patient needs and available medical expertise. Current matching systems often rely on simplistic text matching, failing to capture the nuances of medical inquiries and physician specializations. This results in misallocated resources, delayed responses, and suboptimal patient satisfaction. The paper aims to bridge this gap by developing a more sophisticated system that can accurately and efficiently connect patients with the right experts.
Methods
The core of this research is the Hybrid-GRUCNN-Attention model. This framework employs a dual-channel architecture:
- Patient Problem Information Channel: This channel uses Convolutional Neural Networks (CNNs) to extract deep semantic features from patient question titles, content, personal/disease information, and semantic tags. CNNs are adept at capturing local patterns and hierarchical features within text.
- Medical Expert Information Channel: This channel utilizes Bidirectional Gated Recurrent Units (Bi-GRUs) to process physicians' historical reply sequences, capturing temporal dependencies and evolving expertise. Bi-GRUs are effective for sequential data.
Crucially, an attention mechanism is integrated into both channels. This mechanism allows the model to dynamically weigh the importance of different features, focusing on the most relevant aspects of both the patient's query and the physician's profile. This fine-grained feature interaction is key to improving matching precision. The model fuses these multi-dimensional features to generate a compatibility score between a patient and a physician.
Datasets
The study utilized real-world, de-identified data from two major Chinese healthcare platforms: "YouWenBiDa" and "Xun Yi Wen Yao." The datasets were collected between January 1, 2018, and December 31, 2022.
The patient data included:
- Question titles (average 15 characters)
- Question content (average 128 characters)
- Personal/disease-related details (age, gender, region, disease, medical history, medication records, allergy history)
- Semantic tags (derived from LDA topic modeling)
Physician data included:
- Professional credentials (title, hospital tier, specialty)
- Historical reply sequences (text of answers spanning ≥3 years)
The combined dataset was split into training, validation, and test sets (6:2:2 ratio).
Key Insights and Ideas
- Multimodality is Key: Simply relying on text content is insufficient. Integrating diverse features like question titles, semantic tags, and physician credentials significantly boosts recommendation accuracy.
- Attention is Crucial for Nuance: The attention mechanism is vital for dynamically highlighting critical semantic features, allowing the model to understand the subtle relationships between patient needs and expert capabilities.
- Temporal Dynamics Matter for Experts: Physician expertise isn't static. Bi-GRUs effectively capture the evolution of a physician's knowledge and response patterns over time.
- Balancing Precision and Recall: The model demonstrates an impressive ability to achieve high precision (finding the right expert) while also maintaining good recall (not missing potentially suitable experts), even for less popular or "long-tail" physicians.
- Real-time Performance is Achievable: The proposed model is not only accurate but also efficient, meeting stringent latency requirements for practical clinical deployment.
Key Results
- Accuracy and AUC: The Hybrid-GRUCNN-Attention model achieved a remarkable 89.3% accuracy and 88.0% AUC on the full dataset, significantly outperforming baseline models.
- Precision and Recall: The model achieved Precision@10 of 0.675 and Recall@10 of 0.671, showing substantial improvements over traditional methods and even advanced baselines.
- Real-time Performance: The model boasts an average inference time of 85 ms and a P99 latency of 220 ms, meeting emergency-grade response requirements and enabling efficient clinical deployment.
- Popularity Bias Mitigation: The model effectively balances exposure between popular and niche experts, increasing tail expert exposure by 50% while maintaining high recall for top specialists.
- Feature Importance: Ablation studies showed that integrating multiple input features (title, content, tags, physician credentials, and history) yielded the best performance, with full context integration achieving peak results.
Limitations and Future Work
The authors acknowledge two main limitations:
- Semantic Ambiguity: When patient descriptions are vague or incomplete, the consistency of the system's understanding can decrease.
- Data Sparsity: For newly registered or less active physicians, scarce historical response data can impact the model's effectiveness, although knowledge graph embeddings help mitigate this.
Future work directions include:
- Dynamic Feature Enhancement: Incorporating longitudinal tracking of patient social network dynamics and more comprehensive physician evaluations.
- Cross-Modal Reasoning: Synergizing medical knowledge graph embeddings with other modalities to improve contextual semantic precision.
- Adaptive Transfer Learning: Deploying frameworks for domain-invariant representation learning to address physician-specific data sparsity.
- Real-world Clinical Integration: Collaborating with mainstream medical platforms to refine model performance and facilitate translational applications.
Summary updated