研究目的

本研究旨在探究中国成年女性尿失禁（Urinary Incontinence，UI）的转归情况，分析压力性尿失禁（Stress Urinary Incontinence，SUI）、急迫性尿失禁（Urgency Urinary Incontinence，UUI）和混合性尿失禁（Mixed Urinary Incontinence，MUI）三种亚型的症状缓解、亚型转换及持续存在的变化规律，同时深入挖掘影响中国成年女性UI转归的关键因素。此外，基于机器学习算法，运用logistic回归模型、随机森林模型和贝叶斯网络模型，构建并验证适用于中国成年女性UI转归的预测模型，为临床早期识别高风险人群、制定个体化干预措施、提升UI临床缓解率和改善患者生活质量提供科学依据和技术支持。

研究方法

本研究的研究数据来源于2014年2月至2016年1月实施的“女性盆底功能障碍性疾病流行病学调查”。该调查采用多阶段分层整群抽样法，从中国六大地理区域中随机选取6个代表性省份，纳入年龄≥ 20岁、在当地连续居住至少5年且基线调查时患有UI的女性，排除妊娠期女性及存在严重精神障碍或躯体疾病的女性。数据分析采用IBM SPSS Statistics 26.0和R 4.4.2软件。分类变量数据的特征用频数（百分比）描述，连续性变量数据根据其正态性用平均值±标准差或中位数（四分位数间距）表示。采用Pearson χ²检验比较收集得到的不同变量（如年龄、地区、民族等）对UI缓解、SUI转归、UUI转归及MUI转归的影响，将P < 0.10的变量纳入模型。分别通过logistic回归模型、随机森林模型和贝叶斯网络模型三种机器学习算法构建UI转归预测模型，使用五折交叉验证评估模型性能，评价指标包括ROC曲线下面积（Area Under Curve，AUC）、准确度、灵敏度、特异度等。另外，采用基尼指数平均下降量对随机森林模型进行变量重要性排序；对贝叶斯网络模型进行置换重要性排序。

研究结果

1. 研究共纳入4 695例UI女性患者，中位随访时间4.0年，平均年龄52.6 ± 13.9岁。各UI亚型中，最为常见的是SUI，占比55.8%（2 620 / 4 695）；其次为MUI，占比32.5%（1 524 / 4 695）；数量最少的是UUI，占比11.7 %（551 / 4 695）。UI 4年缓解率为49.0%，年均缓解率为12.3%，SUI年均缓解率为12.6%，UUI年均缓解率为13.3%，MUI年均缓解率为11.3%。

2. 平均每年有0.3%的SUI患者转化为UUI，4.0%的SUI患者转化为MUI；平均每年有7.6%的UUI患者转化为SUI，8.6%的UUI患者转化为MUI；平均每年有7.7%的MUI患者转化为SUI，1.8%的MUI患者转化为UUI。

3. 年龄增长、肥胖、慢性咳嗽及糖尿病等因素是UI症状持续的关键危险因素。

4. 在UI缓解模型中贝叶斯网络模型表现出最佳性能，AUC为0.669；在各亚型转归预测模型中随机森林模型展现出最优的综合性能，AUC值均超过0.7，最高达0.976。

研究结论

1. UI各亚型转归情况具有显著的差异，其中UUI患者4年缓解率最高，SUI次之，MUI最低。

2. 研究进一步证实，年龄增长、肥胖、慢性咳嗽及糖尿病等因素是UI症状持续的关键危险因素。

3. 在构建预测模型的过程中，随机森林算法在多种亚型转归预测中展现出最优的综合性能，其在处理非线性关系和高维数据方面的能力，为复杂转归模式的解析提供了有效的工具；logistic回归模型则凭借其显著的临床可解释性，更适合于基层医疗场景中的初步风险评估。

Objective
This study aimed to investigate the prognosis of urinary incontinence (UI) among adult women in China by analyzing the dynamic patterns of symptom remission, subtype transitions, and persistence across three UI subtypes: stress urinary incontinence (SUI), urgency urinary incontinence (UUI), and mixed urinary incontinence (MUI). Additionally, the study sought to identify key factors influencing UI prognosis and to develop and validate machine learning prediction models (logistic regression model, random forest model, and Bayesian network model) for clinical applications, enabling early identification of high-risk populations, personalized interventions, and improved clinical remission rates and quality of life for UI patients.
Methods
Data were derived from the "Epidemiological Survey on Female Pelvic Floor Dysfunction" conducted between February 2014 and January 2016. A multi-stage stratified cluster sampling method was employed to recruit participants from six representative provinces across China’s six major geographic regions. Eligible participants included women aged ≥ 20 years, residing locally for ≥ 5 consecutive years, and diagnosed with UI at baseline, excluding pregnant women and those with severe mental or physical disorders. Statistical analyses were conducted using IBM SPSS Statistics 26.0 and R 4.4.2 software. Categorical variables were expressed as frequencies with percentages, while continuous variables were presented as mean ± standard deviation for normally distributed data or median with interquartile range (IQR) for non-normally distributed variables. Pearson’s χ2 test was employed to analyze associations between demographic characteristics (including age, geographic region, and ethnic background) and UI outcome measures: remission status, and transitions between SUI, UUI, and MUI subtypes. Variables demonstrating univariate associations at P < 0.10 were retained for multivariable modeling Three machine learning algorithms (logistic regression, random forest, and Bayesian network models) were utilized to construct UI prognosis prediction models. Model performance was evaluated via five-fold cross-validation, with metrics including the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. Variable importance was ranked using the mean decrease in Gini index for the random forest model and permutation importance for the Bayesian network model.
Results
1. The cohort comprised 4,695 women diagnosed with UI, with median follow-up of 4.0 years. Participants had a mean baseline age of 52.6 years (SD: 13.9). SUI demonstrated the highest prevalence (n = 2,620; 55.8%), followed by MUI (n = 1,524; 32.5%), with UUI constituting the smallest proportion (n = 551; 11.7%). The four-year UI remission rate was 49.0%, with an annual remission rate of 12.3%. Subtype-specific annual remission rates were 12.6% for SUI, 13.3% for UUI, and 11.3% for MUI. 
2. Transition analysis revealed that SUI patients exhibited an annual transition rate of 0.3% to UUI and 4.0% to MUI; UUI patients transitioned at 7.6% to SUI and 8.6% to MUI annually; and MUI patients transitioned at 7.7% to SUI and 1.8% to UUI annually. 
3. Key risk factors for UI persistence included aging, obesity, chronic cough, and diabetes. 
4. Among prediction models, the Bayesian network model demonstrated superior performance for UI remission prediction (AUC: 0.669), while the random forest model achieved the highest overall performance for subtype-specific prognosis prediction, with AUC values exceeding 0.7 (up to 0.976).
Conclusion
1. UI subtypes exhibited significant disparities in clinical outcomes. UUI demonstrated the highest 4-year remission rate, followed by SUI, whereas MUI showed the lowest remission probability. 
2. Aging, obesity, chronic cough, and diabetes were confirmed as critical determinants of UI persistence. 
The random forest algorithm excelled in predicting complex subtype transitions due to its capacity to handle nonlinear relationships and high-dimensional data, while the logistic regression model, with its clinical interpretability, proved advantageous for preliminary risk assessment in primary healthcare settings.