Developing a comprehensive risk prediction model for the emergence of lower extremity lymphedema following gynecological cancer surgery by harnessing the power of integrative multi-factor analysis and cutting-edge machine learning algorithms

Abstract Objective Establish the predictive efficacy of postoperative lower extremity lymphedema in gynecological malignant tumors to precisely identify high-risk individuals and enable timely, proactive interventions. Methods A comprehensive retrospective analysis was conducted on the clinical records of 712 patients with gynecological malignant tumors who underwent surgical intervention at a tertiary hospital in China between January 2016 and January 2021. The postoperative disease trajectories were meticulously documented and evaluated. Based on the development of lower extremity lymphedema, patients were stratified into an occurrence group (n = 121) and a non-occurrence group (n = 591). Data were randomly allocated into training and validation sets using a 7:3 partition ratio, with the latter comprising 30% of the total cohort. The training set served as the foundation for model development, while the validation set ensured rigorous assessment of predictive performance. Through logistic regression analysis, a refined set of significant predictors was identified. Subsequently, a robust risk prediction model for post-surgical lower extremity lymphedema was constructed by integrating traditional logistic regression with five advanced machine learning algorithms—random forest, support vector machine, gradient boosting machine, neural network, and extreme gradient boosting—thereby harnessing both statistical rigor and computational intelligence to enhance predictive accuracy. Results Univariate logistic analysis unveiled that tumor type, FIGO stage, pathological grade, number of pelvic lymph node dissections, postoperative radiotherapy, hypertension, and prolonged postoperative daily standing time were significantly associated with an elevated risk of lower extremity lymphedema in patients with gynecological malignancies, whereas increased postoperative daily exercise duration emerged as a protective determinant. Multivariate logistic regression further identified hypertension, advanced FIGO stage, extensive pelvic lymph node resection, postoperative radiotherapy, and prolonged standing time as independent predictors of disease onset. Leveraging traditional logistic regression alongside five widely adopted machine learning algorithms, a comprehensive predictive model was meticulously constructed and rigorously evaluated. Through systematic screening across three critical dimensions—discriminative accuracy, calibration fidelity, and clinical utility—the gradient boosting machine (GBM) demonstrated superior performance, outperforming all competing models. SHapley Additive exPlanations (SHAP) analysis elucidated the hierarchical importance of predictive features: FIGO stage ranked highest in influence, followed by postoperative daily standing time, number of pelvic lymph node resections, and postoperative radiotherapy, with hypertension contributing the least. Finally, through in-depth feature effect profiling, dependency structure exploration, and individualized prediction interpretation, the model’s behavioral patterns were validated to align seamlessly with the findings of multivariate logistic regression, affirming both its statistical consistency and clinical interpretability. Conclusion The gradient boosting machine (GBM) prediction model, meticulously constructed by integrating five pivotal factors—hypertension, FIGO stage, extent of pelvic lymph node resection, postoperative radiotherapy, and daily standing duration following surgery—demonstrates robust predictive performance in forecasting the onset of lower extremity lymphedema after gynecological malignant tumor surgery. Clinically, this evidence-based risk assessment tool empowers healthcare providers to stratify patient risk with greater precision, offering a valuable framework for clinical consultation, identification of high-risk cohorts, and personalized follow-up strategies, thereby facilitating early intervention and optimized therapeutic decision-making.摘要 目的 构建妇科恶性肿瘤术后下肢淋巴水肿的预测效能评估体系，以精准甄别高危人群并实施及时主动的临床干预。 方法 对2016年1月至2021年1月期间国内某三级医院收治的712例接受手术治疗的妇科恶性肿瘤患者的临床病历开展回顾性综合分析。详细记录并评估患者术后疾病转归情况，依据是否发生下肢淋巴水肿将患者分为发生组（n=121）与未发生组（n=591）。采用7:3的划分比例将数据随机划分为训练集与验证集，其中验证集占总队列的30%。训练集用于模型构建，验证集用于严格评估预测性能。通过logistic回归（logistic regression）分析筛选出具有统计学意义的预测因子，随后将传统logistic回归与五种先进机器学习算法——随机森林（random forest）、支持向量机（support vector machine）、梯度提升机（gradient boosting machine, GBM）、神经网络（neural network）以及极端梯度提升（extreme gradient boosting, XGBoost）相结合，构建稳健的术后下肢淋巴水肿风险预测模型，兼顾统计严谨性与计算智能以提升预测精度。 结果 单因素logistic回归分析显示，肿瘤类型、国际妇产科联盟（International Federation of Gynecology and Obstetrics, FIGO）分期、病理分级、盆腔淋巴结清扫数目、术后放疗情况、高血压病史以及术后每日站立时长与妇科恶性肿瘤患者术后下肢淋巴水肿的发病风险显著相关；而术后每日运动时长则为保护性因素。多因素logistic回归进一步明确，高血压病史、晚期FIGO分期、大范围盆腔淋巴结清扫、术后放疗以及延长的站立时长为该病发生的独立预测因子。本研究构建并严格评估了结合传统logistic回归与五种主流机器学习算法的综合预测模型。通过判别准确性、校准度以及临床实用性三个核心维度的系统筛查，梯度提升机（GBM）展现出最优性能，优于其余所有对比模型。SHAP可加性解释（SHapley Additive exPlanations, SHAP）分析阐明了预测特征的层级重要性：FIGO分期的影响程度最高，其次为术后每日站立时长、盆腔淋巴结清扫数目以及术后放疗，高血压病史的贡献度最低。最后，通过深入的特征效应分析、依赖结构探索以及个体化预测解读，验证了模型的行为模式与多因素logistic回归的结果高度一致，证实了其统计一致性与临床可解释性。 结论 本研究构建的梯度提升机（GBM）预测模型，整合了高血压病史、FIGO分期、盆腔淋巴结清扫范围、术后放疗以及术后每日站立时长五项关键影响因素，在预测妇科恶性肿瘤术后下肢淋巴水肿发生方面展现出稳健的预测性能。该循证风险评估工具可帮助临床医护人员更精准地进行患者风险分层，为临床会诊、高危人群甄别以及个体化随访策略提供了极具价值的框架，从而助力早期干预与优化的治疗决策制定。