AutoPCOS: a stepwise multimodal intelligent framework for polycystic ovary syndrome risk stratification and...

Polycystic ovary syndrome (PCOS) is a prevalent endocrine disorder among women of reproductive age, typically diagnosed through a combination of clinical evaluation, laboratory testing, and ultrasonography. However, this multimodal diagnostic pathway is often time-consuming, costly, and dependent on resource availability, thereby limiting its accessibility in real-world clinical settings.

In this study, we propose AutoPCOS, a stepwise multimodal intelligent framework for flexible PCOS risk stratification and diagnostic support. Using a publicly available Kaggle PCOS dataset, features were categorized into three modalities: clinical, laboratory, and ultrasound data. Based on data availability, four predictive models were constructed: (1) clinical-only, (2) clinical + laboratory, (3) clinical + ultrasound, and (4) full multimodal models. Random Forest was employed as the primary classifier, with comparisons against Logistic Regression, Support Vector Machine, Decision Tree, and Gradient Boosting. Subgroup analyses were conducted based on body mass index (BMI) and menstrual cycle patterns.

The proposed framework demonstrated robust predictive performance across different data availability scenarios. Notably, the models achieved strong performance in subgroups with BMI < 24 and irregular menstrual cycles, with precision values reaching ≥ 0.929. Comparative analysis confirmed the effectiveness of the Random Forest model. Furthermore, the integration of a knowledge base and the Lingshu large language model enabled interpretable risk explanations and personalized recommendations.

AutoPCOS provides a flexible and resource-aware framework for PCOS risk assessment that adapts to varying clinical conditions and data accessibility. By supporting stepwise decision-making and enhancing interpretability, the system shows potential as a practical tool for both patients and healthcare providers. Future work will focus on validation using real-world clinical datasets and improving model generalizability.