DocSmile: Innovative Practices of a Series of Large Language Models Specialized in Dentistry

DocSmile is a series of advanced large language models specifically designed for dental intelligent assistance. Through domain-specific fine-tuning technology, it addresses the limitations of general-purpose large language models in dental professional scenarios and provides professional AI solutions for stomatology. It covers multiple scenarios such as clinical decision support, medical education and training, patient services and health management. It emphasizes safety and quality control, adopts an open-source strategy to promote community collaboration, and plans future development directions such as multimodal expansion.

Project Background

Today, as artificial intelligence penetrates various industries, general-purpose large language models lack domain knowledge, making it difficult to understand dental professional terminology and clinical contexts, and unable to provide standard-compliant treatment recommendations. This pain point gave birth to the DocSmile project, whose goal is to create an AI assistant that understands stomatology.

Unique Challenges of Dental AI

1. Highly Visual Diagnosis: Relies on image data such as oral endoscopes, X-rays, CBCT, etc., and requires comprehensive judgment combining symptoms and medical history;
2. Precise Treatment Planning: Needs to understand the indications, contraindications, etc., of treatment plans;
3. Interdisciplinary Knowledge Integration: Covers sub-fields such as maxillofacial surgery and periodontology, requiring interdisciplinary referral recommendations;
4. Patient Communication and Anxiety Management: Needs to have empathetic communication skills to relieve patients' tension.

Domain-Specific Pre-training

Pre-trained on a large amount of corpus including dental professional literature, textbooks, clinical guidelines, de-identified case reports, etc., to learn professional terminology, knowledge systems, and reasoning patterns.

Supervised Fine-tuning and Instruction Alignment

Fine-tuned using instruction datasets such as doctor-patient question-answer pairs, diagnostic reasoning chains, treatment plan comparisons, etc., to improve the model's applicability in real scenarios.

Multimodal Capability Expansion

The current version focuses on text capabilities, with the architecture reserving multimodal interfaces. Future plans include integrating oral image understanding capabilities.

Clinical Decision Support

Provides doctors with support such as differential diagnosis, treatment plan recommendations, drug interaction checks, referral suggestions, etc.

Medical Education and Training

Provides students with learning resources such as virtual case discussions, knowledge Q&A, literature guidance, exam preparation, etc.

Patient Services and Health Management

Helps patients with symptom self-check (non-diagnostic), pre-treatment consultation, post-operative care guidance, oral health education, etc.

Medical Accuracy Assurance

Ensures output quality through retrieval-augmented generation (RAG) technology to link authoritative knowledge bases, expert review feedback, continuous knowledge updates, etc.

Clear Responsibility Boundaries

AI only provides information support and does not replace doctors' judgments. Recommendations need to be reviewed by physicians, and in emergency situations, patients are guided to seek face-to-face help.

Privacy and Data Security

Supports local deployment, end-to-end encryption, strict access control, and audit logs to protect patients' sensitive information.

DocSmile adopts an open-source strategy and is hosted on GitHub. Its values include:

1. Transparency and Trustworthiness: Allows review of implementation details to verify safety and reliability;
2. Collaborative Innovation: Brings together global experts to submit data, improve architecture, develop functions, etc.;
3. Knowledge Sharing: Promotes knowledge sharing in the dental AI field, lowers entry barriers, and accelerates industry development.

Future Development Plans

• Short-term: Expand training data, optimize multilingual support, and improve reasoning capabilities in complex scenarios;
• Mid-term: Integrate oral image understanding, develop personalized treatment plans, and conduct clinical validation;
• Long-term: Build a dental AI ecosystem, promote the establishment of industry standards, and advance global oral health equity.

Conclusion

DocSmile proves that vertical domain large language models can be reliable assistants in professional medical scenarios, which is expected to improve diagnosis and treatment efficiency, enhance patient experience, and promote the improvement of global oral health levels.