| 范文轩,张经纬,李慧武.基于可解释机器学习的全关节置换术后机械性并发症预测建模与风险表型聚类.骨科,2026,17(1): 9-16. |
| 基于可解释机器学习的全关节置换术后机械性并发症预测建模与风险表型聚类 |
| Predictive modeling and risk phenotype clustering for mechanical complications after total joint arthroplasty based on interpretable machine learning |
| 投稿时间：2025-09-01 |
| DOI：10.3969/j.issn.1674-8573.2026.01.003 |
| 中文关键词: 全关节置换术 机械性并发症 可解释机器学习 沙普利加和解释法 风险表型聚类 精准医学 |
| 英文关键词: Total joint arthroplasty Mechanical complications Interpretable machine learning SHAP Risk phenotype clustering Precision medicine |
| 基金项目:国家自然科学基金项目（82472476） |
|
| 摘要点击次数: 15 |
| 全文下载次数: 0 |
| 中文摘要: |
| 目的 本研究旨在构建基于可解释机器学习的全关节置换术（total joint arthroplasty，TJA）术后机械性并发症预测模型，并探索潜在的风险表型异质性。方法 基于英国生物银行（UK Biobank）队列，采用1∶3病例-对照匹配设计。利用自动化机器学习（AutoML）框架构建包含XGBoost、CatBoost、LightGBM等多算法的集成预测模型，采用分层5折交叉验证进行训练和优化。运用沙普利加和解释法（SHapley Additive exPlanations，SHAP）进行模型可解释性分析，识别关键风险因素及其贡献模式。基于个体SHAP值向量进行无监督聚类分析，统一流形逼近与投影（uniform manifold approximation and projection，UMAP）降维技术识别不同的风险表型亚组。结果 最优集成模型在测试集上获得曲线下面积（AUC）为0.773的预测性能，前10个候选模型的AUC集中在0.769~0.773的窄区间内。SHAP分析显示，残障蓝徽章资格、身体质量指数（BMI）、步行速度、尿素、维生素D等为主要风险驱动因素。基于SHAP值聚类识别出15个病人亚组，机械并发症发生率呈显著梯度分布（7.7%~71.4%），相对风险比范围为0.31~2.86。进一步归纳为六类主导风险模式：代谢/体成分相关型（6簇）、运动功能受限/虚弱型（4簇）、肾功能相关型（2簇）、炎症/免疫相关型（1簇）、社会经济/生活方式型（1簇）和营养/微量元素相关型（1簇）。结论 基于可解释机器学习的集成模型能够有效预测TJA术后机械性并发症风险，SHAP驱动的风险表型聚类揭示了病人风险异质性的内在结构。六类风险模式为个体化术前评估、围术期优化和精准干预提供了可操作的临床决策框架，有望推动TJA并发症精准预防的临床转化应用。 |
| 英文摘要: |
| Objective To construct an interpretable machine learning-based prediction model for mechanical complications after total joint arthroplasty (TJA) and explore potential risk phenotype heterogeneity. Methods Based on the UK Biobank cohort, a 1∶3 case-control matched design was employed. An ensemble prediction model incorporating multiple algorithms including XGBoost, CatBoost, and LightGBM was constructed using an automated machine learning (AutoML) framework, with stratified 5-fold cross-validation for training and optimization. SHapley Additive exPlanations (SHAP) methodology was utilized for model interpretability analysis to identify key risk factors and their contribution patterns. Unsupervised clustering analysis based on individual SHAP value vectors, combined with Uniform manifold approximation and projection (UMAP) dimensionality reduction techniques, was performed to identify different risk phenotype subgroups. Results The optimal ensemble model achieved a predictive performance of AUC=0.773 on the test set, with the top 10 candidate models all concentrated within a narrow AUC range of 0.769-0.773. SHAP analysis revealed that disability blue badge eligibility, body mass index (BMI), walking speed, urea, and vitamin D were the primary risk-driving factors. SHAP value-based clustering identified 15 patient subgroups with mechanical complication rates showing significant gradient distribution (7.7%-71.4%) and relative risk ratios ranging from 0.31-2.86. These were further categorized into six dominant risk patterns: metabolic/body composition-related (6 clusters), motor function limitation/frailty-related (4 clusters), renal function-related (2 clusters), inflammation/immune-related (1 cluster), socioeconomic/lifestyle-related (1 cluster), and nutrition/micronutrient-related (1 cluster). Conclusion The interpretable machine learning-based ensemble model can effectively predict mechanical complication risks after TJA, and SHAP-driven risk phenotype clustering reveals the intrinsic structure of patient risk heterogeneity. The six risk patterns provide an actionable clinical decision-making framework for individualized preoperative assessment, perioperative optimization, and precision intervention, potentially promoting clinical translational applications for precision prevention of TJA complications. |
|
查看全文
下载PDF阅读器 |
| 关闭 |
|
|
|
