Efficacy prediction of acupuncture and moxibustion combined with conventional treatment for PCOS-IR based on real-world data: machine learning and interpretability analysis

To construct an efficacy prediction model for polycystic ovary syndrome with insulin resistance (PCOS-IR) treated with acupuncture and moxibustion combined with conventional treatment based on machine learning.

Data from two real-world studies were collected for the training set (284 cases) and validation set (132 cases). The training set included the data of PCOS-IR patients visited from January 2023 to September 2024, and the validation set was composed of the patients visited from September 2024 to February 2025. Logistic regression (LR) and random forest (RF) algorithms were combined for predictive feature selection, and 5 representative machine learning models with different principles were built based on the screening results. The predictive effectiveness of the best model was assessed by receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis (DCA). Finally, the predictive results of the best model were interpreted using the Shapley additive interpretation (SHAP) framework.

The predictive features used for model construction included fasting insulin (FINS), total cholesterol (TC), the upper limit of the menstrual cycle length cycle (UML), body mass index (BMI), and alanine aminotransferase (ALT). The RF model showed the most balanced performance in terms of accuracy, precision, F1 score, and area under curve (AUC), suggesting that it achieved the overall favorable performance in predicting the efficacy on PCOS-IR. The SHAP analysis further revealed the importance of these 5 predictive features in efficacy prediction; and in particular, FINS, as an indicator of insulin level, was conductive most significantly to the efficacy prediction.

The efficacy prediction model constructed for PCOS-IR treated with acupuncture and moxibustion, combined with conventional regimens, provides an important empirical evidence for identifying beneficiary population and optimizing treatment strategy.