@Female Comrades Your AI Chief Physician Is Here! China’s First Gynecology Large Model in a Vertical Domain Passes the “Senior Professional Title” Exam

On July 28, the "Doukou Gynecology Large Model," developed by Yisheng Jiankang (Hangzhou) Life Science Technology Co., Ltd., successfully passed the written exam for the National Senior Professional Title in Obstetrics and Gynecology (Senior Level). This makes it the first vertical medical model in China trained on DingTalk's enterprise-specific AI platform to meet this professional standard. This breakthrough not only marks a key leap in the development of large models in the medical vertical domain but also provides a replicable innovation path for the industry with a clear demonstration effect.

Technical Breakthroughs and Efficient R&D Path

From technical攻坚 to practical implementation, the Doukou Gynecology Large Model completed its development and training—from zero to excellence—in a short period of time, achieving professional qualification standards. This highly efficient breakthrough challenges the conventional belief that developing vertical-domain large models requires high investment and long timelines, proving that small and medium-sized teams can, with the support of specialized training platforms, rapidly build proprietary large models that reach top professional levels through scientific training methods, high-quality domain data, and focused technical efforts.

Core Technical Support and Training Methods

The core technical support behind this breakthrough is an advanced foundation model. Leveraging DingTalk's enterprise-specific AI platform and professional services, the Doukou Gynecology Large Model achieved performance leaps through high-quality obstetrics and gynecology data construction and multi-stage optimization training methods.

Zhu Hong, CTO of DingTalk, noted that the Doukou Gynecology Large Model is the first specialized vertical large model born on the DingTalk AI platform. In just over a month of collaboration between the two teams, the model's accuracy was boosted to 90.2%, and it successfully passed the professional exam. This validates DingTalk's ability to help enterprises across industries build proprietary large models. "DingTalk is continuously improving its support system for building industry- and enterprise-specific large models, creating a pay-per-performance model for AI large models to help more industry enterprises like Yisheng Jiankang truly implement AI applications," Zhu Hong said.

In the development process of the Doukou Gynecology Large Model, Yisheng Jiankang and DingTalk adopted a technical approach that combines "high-quality, precisely annotated medical data + customized training tools + efficient training processes and methods." This enabled rapid iteration of the model while significantly enhancing its accuracy and stability, allowing it to perform exceptionally well in complex clinical scenarios. The practices employed in the Doukou Gynecology Large Model—from data preparation and preprocessing to continuous performance optimization—provide a replicable reference case for building specialized large models in the medical field and beyond.

Exam Standards and Evaluation Results

The National Senior Professional Title (Senior Level) exam in Obstetrics and Gynecology is the gold standard for measuring the professional competence of obstetricians and gynecologists. The exam covers 12 core subject areas, including female reproductive system anatomy, clinical obstetrics and gynecology, and reproductive endocrinology, and places particular emphasis on practical skills such as diagnosing complex cases and designing high-difficulty surgical plans, requiring candidates to possess "clinical intuition" built up over decades of clinical experience.

The written exam evaluation strictly followed the People's Medical Publishing House's "Full-Scale Mock Exam for Senior Obstetrics and Gynecology Titles," as specified by the National Health Commission. The exam scope covered 12 core subjects, including clinical obstetrics and gynecology, gynecologic oncology, perinatal medicine, reproductive endocrinology, and family planning. The question types included multiple-choice questions (40% of the total) and case-analysis questions (60%). The case-analysis questions required the model to address clinical diagnosis, differential diagnosis, and treatment plans based on patient complaints, examination reports, and other multi-source information, comprehensively testing clinical decision-making ability. Only fully correct answers earned points, and the scoring criteria were stricter than those used in actual human exams. Multiple-choice accuracy: 75.56%; case-analysis (multiple-answer) accuracy: 59.01%; overall accuracy: 64.94%. Both the multiple-choice and case-analysis accuracy rates outperformed several other models. To ensure result reliability, the team used a verification method that averaged scores from three independent test papers.

(Compared with models tested on the same exam paper)

Application Prospects and Industry Impact