Domain-Specific GPT Model Enhances Clinical Support in Cleft Lip and Palate

Overview

A specialized language model, CLP-GPT, was developed using parameter-efficient fine-tuning on expert-verified data for cleft lip and palate (CLP). It demonstrated superior accuracy in physician-oriented tasks and comparable performance to leading models in patient communication, offering an efficient clinical decision support tool.

Background

Cleft lip and palate (CLP) is a common congenital anomaly requiring multidisciplinary lifelong management. Despite established clinical protocols, access to specialized knowledge remains limited for patients and primary care providers. Large language models (LLMs) have potential to bridge these gaps, but generalist models often suffer from hallucinations and high computational demands. This study aimed to create a domain-specific, parameter-efficient LLM tailored for CLP clinical applications.

Data Highlights

Key Findings

• CLP-GPT significantly outperformed Claude-3.5-Sonnet and Gemini-1.5-Pro in physician-oriented accuracy (p=0.018 and p<0.001 respectively).
• CLP-GPT's accuracy was nominally higher than GPT-4o but without statistical significance.
• For patient-oriented questions, CLP-GPT scored highly in layperson-rated credibility (4.66) and comprehensibility (4.46), comparable to GPT-4o and Claude-3.5-Sonnet.
• Expert clinicians confirmed CLP-GPT's responses were accurate and complete for patient queries.
• The model was fine-tuned using a rigorously curated dataset of 7,815 expert-verified Q&A pairs derived from public datasets and peer-reviewed literature.
• Subjective evaluations used a 5-point Likert scale with consensus arbitration to minimize bias.

Clinical Implications

CLP-GPT offers a reliable, cost-effective tool for clinical decision support and patient education in cleft lip and palate management. Its parameter-efficient design enables deployment in resource-limited settings while maintaining high accuracy and readability. This model can help bridge knowledge gaps for both clinicians and patients, potentially improving care quality and accessibility.

Conclusion

The development of CLP-GPT demonstrates that specialized, parameter-efficient language models can achieve expert-level clinical accuracy and patient communication quality in cleft lip and palate care. This approach may serve as a scalable solution for domain-specific AI integration in healthcare.

References

1. Huang et al. 2024 -- A Domain-Specific Language Model for Clinical Applications in Cleft Lip and Palate