基于机器学习的患者特异性Fontan失败风险计算器开发
该研究利用澳大利亚和新西兰Fontan登记处(ANZFR)1975至2023年的数据,采用正则化Cox回归从54个术前预测因子中筛选变量,最终基于6个预测因子构建了术前Fontan失败风险计算器,并在独立队列中进行了外部验证。
要点
- 基于6个术前预测因子构建了简易的Fontan失败风险预测模型。
- 模型在独立队列外部验证中表现可接受。
- 计算器已实现为桌面应用,辅助术前个体化咨询。
核心结论
- 机器学习可应用于两国Fontan登记处的大数据,开发术前患者特异性Fontan失败风险计算器。
研究设计与样本
- 数据来自澳大利亚和新西兰Fontan登记处(ANZFR),共纳入1473例患者构建计算器。
- 中位Fontan手术年龄4.5岁(四分位距3.7-5.6岁),中位随访11.0年(四分位距5.3-17.8年)。
- 先使用正则化Cox回归从54个术前预测因子中自动筛选变量,再手动加入临床相关预测因子,最终用6个预测因子进行多变量Cox回归。
主要发现
- 总体队列10年、20年、30年无Fontan失败生存率分别为92%(CI 90%-93%)、83%(CI 80%-86%)和72%(CI 65%-78%)。
- 最终模型使用的6个预测因子为:年龄、心室形态、原发诊断、完全性肺静脉异位引流、Fontan类型、中度或以上房室瓣反流。
- 外部验证在独立队列中显示模型性能可接受。
临床怎么用
- 计算器已用R语言的Shiny库实现为桌面应用,可基于6个术前预测因子进行个体化风险预测。
- 模型可用于患者特异性术前咨询和设定现实期望。
局限性
- 模型将随着更多数据的加入而继续学习和改进。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: The Fontan operation is the final step in staged palliation for patients with single-ventricle physiology. It has extended their life expectancy and improved their quality of life. However, long-term complications and Fontan failure remain lifelong concerns. We aimed to use machine learning to develop a patient-specific preoperative Fontan failure risk calculator. METHODS: Patient data were obtained from the Australia and New Zealand Fontan Registry (ANZFR). The primary composite end point was Fontan failure, defined as any of death, transplant, Fontan takedown or conversion, protein-losing enteropathy, plastic bronchitis, or New York Heart Association class III/IV. To construct the risk calculator, we first used Cox regression with regularization to predict Fontan failure from 54 preoperative predictors in the ANZFR database. A regularization machine learning tool was used to automate variable selection among many predictors. We then manually added clinically relevant predictors. Six predictors (age, ventricular morphology, primary diagnosis, total anomalous pulmonary venous drainage, Fontan type, and moderate or greater atrioventricular valve regurgitation) were ultimately used in a subsequent multivariable Cox regression (without regularization) to ensure the final risk prediction model was simple and easy to interpret. RESULTS: Data from 1888 patients over 48 years (1975-2023) were available. The ANZFR collects perioperative and follow-up variables about each patient. After excluding patients with Fontan procedures with an atriopulmonary connection (n = 290) and missing predictors or outcome data (n = 125), data from 1473 patients were used to construct the calculator. Median age at Fontan was 4.5 years (interquartile range, 3.7, 5.6 years). Median follow-up was 11.0 years (interquartile range, 5.3, 17.8 years). Freedom from Fontan failure for the overall cohort at 10, 20, and 30 years was 92% (confidence interval [CI], 90%-93%), 83% (CI, 80%-86%), and 72% (CI, 65%-78%), respectively. External validation in an independent cohort demonstrated acceptable model performance. The risk prediction model was then implemented in a Desktop application using the Shiny library in R and used to develop the preoperative Fontan failure calculator on the basis of the 6 predictors. CONCLUSIONS: Machine learning can be applied to "big data" from a binational Fontan Registry to develop a preoperative, patient-specific Fontan failure risk calculator. The model will continue to learn and improve as more data is added. This is a step toward personalized medicine enabling patient-specific pre-operative counselling and realistic expectations.
引用
- Marathe SP, Betts KS, Venna A, et al. Development of a patient-specific Fontan failure risk calculator using machine learning-a step toward personalized medicine. J Thorac Cardiovasc Surg. 2026. PMID: 41548840. doi:10.1016/j.jtcvs.2025.12.032
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。