Abstract
INTRODUCTION AND HYPOTHESIS: We aimed to develop and validate a clinically applicable risk assessment model for identifying women at a high risk of pelvic organ prolapse (POP) based on a retrospective practice. METHODS: This study enrolled patients with and without POP between January 2019 and December 2021. Clinical data were collected and machine learning models were applied, such as multilayer perceptron, logistic regression, random forest (RF), light gradient boosting machine and extreme gradient boosting. Two datasets were constructed, one comprising all variables and the other excluding physical examination variables. Two versions of the machine learning model were developed. One was for professional doctors, and the other was for community-health providers. The area under the curve (AUC) and its confidence interval (CI), accuracy, F1 score, sensitivity, and specificity were calculated to evaluate the model's performance. The Shapley Additive Explanations method was used to visualize and interpret the model output. RESULTS: A total of 16,416 women were recruited, with 8,314 and 8,102 in the POP and non-POP groups respectively. Eighty-seven variables were recorded. Among all candidate models, the RF model with 13 variables showed the best performance, with an AUC of 0.806 (95% CI 0.793-0.817), accuracy of 0.723, F1 of 0.731, sensitivity of 0.742, and specificity of 0.703. Excluding the physical examination variables, the RF model with 11 variables showed an AUC, accuracy, F1 score, sensitivity, and specificity of 0.716, 0.652, 0.688, 0.757, and 0.545 respectively. CONCLUSIONS: We constructed a clinically applicable risk warning system that will help clinicians to identify women at a high risk of POP.