基于可解释机器学习算法的产后压力性尿失禁预测模型研究

关键词：产后康复  疾病预测模型  可解释机器学习  SHAP  压力性尿失禁

Keywords:postpartum  disease prediction model  explainable machine learning  SHAP  stress urinary incontinence

摘要：

      摘要 目的：使用多临床因素结合机器学习算法构建产后压力性尿失禁（stress urinary incontinence，SUI）的疾病预测模型，并使用SHAP（Shapley additive explanations）算法解释最优模型中各因子在决策过程中的影响方向及贡献。 方法：回顾性分析2021年4月至2024年9月于徐州市中心医院新城分院产后康复科就诊的产后复查患者资料3114例。收集产妇临床信息，使用秩和检验和χ2检验筛选健康组和患病组间具有差异性的特征因子后，使用最小绝对收缩和选择算子（least absolute shrinkage and selection operator， LASSO） 和Boruta算法进行二次特征选择。基于所选特征构建10种常见机器学习模型并比较其最终预测效果。通过计算模型的曲线下面积（area under the curve，AUC）、准确率、精确度、召回率和F1指数评估算法效能，选择最优模型进行SHAP解释。 结果：随机梯度提升（stochastic gradient boosting，SGboost）算法模型的AUC 0.804、准确率0.855、精确度0.861、召回率0.978、F1分数0.916，预测性能最优。SHAP算法解析表明，阴道前壁中度膨出、年龄、顺产、胎重、SUI家族史、身体质量指数、Ⅰ类肌收缩值以及耐力收缩幅度在SUI发生中为关键预测因子，累计权重82.3%。 结论：基于多维度临床特征的SGBoost模型在产后SUI的早期风险分层中具有可行性，但其临床应用仍需多中心外部数据验证。

Abstract:

      Abstract Objective： To construct a disease prediction model for postpartum stress urinary incontinence（SUI） using multiple clinical factors combined with machine learning algorithms， and to interpret the direction and contribution of each factor in the decision-making process of the optimal model using the Shapley additive explanations（SHAP） algorithm. Method： A retrospective analysis was conducted on 3，114 postpartum patients who visited the Postpartum Rehabilitation Department of the Xuzhou Central Hospital from April 2021 to September 2024. Clinical information of the patients was collected， and characteristic factors with differences between the healthy and diseased groups were screened using the Mann-Whitney U test and chi-square test. Least absolute shrinkage and selection operator（LASSO） and Boruta algorithms were then used for secondary feature selection. Based on the selected features， ten common machine learning models were constructed， and their final prediction performance was compared. The performance of the algorithms was evaluated by calculating the AUC（area under the curve），accuracy，precision，recall， and F1 score of the models， and the optimal model was selected for SHAP interpretation. Result： The stochastic gradient boosting （SGBoost） algorithm model achieved an AUC of 0.804， accuracy of 0.855， precision of 0.861， recall of 0.978， and F1 score of 0.916， with the highest prediction performance. SHAP analysis revealed that moderate anterior vaginal wall prolapse， age， vaginal delivery， fetal weight， family history of SUI， BMI， type I muscle fiber contraction value， and endurance contraction amplitude were the key predictors of SUI， collectively accounting for 82.3% of the model’s predictive weight. Conclusion： The SGBoost model incorporating multidimensional clinical features demonstrates potential for early risk stratification of postpartum SUI. However， its clinical applicability necessitates further validation through multi-center external datasets to ensure generalizability.