临床医生与计算方法预测间歇性跛行不良结局的性能比较
本研究基于99例间歇性跛行患者的匿名数据集,比较机器学习、传统逻辑回归与8名临床医生对不良结局的预测性能。结果显示,基于LASSO的机器学习模型在所有结局上的受试者工作特征曲线下面积、F1分数和Brier分数均显著优于逻辑回归和临床医生。
要点
- 机器学习模型预测2年慢性肢体威胁性缺血(CLTI)的AUC为0.885,逻辑回归为0.74,最佳临床医生为0.63。
- 机器学习模型预测5年主要不良心血管事件的AUC为0.963,逻辑回归为0.759,最佳临床医生为0.611。
- 机器学习模型Brier分数为0.03-0.07,逻辑回归为0.10-0.22,临床医生大于0.31。
核心结论
- 基于机器学习的预测模型显著优于传统回归和临床医生判断,主要因其能捕捉变量间复杂的非线性关联。
研究设计与样本
- 研究使用99例间歇性跛行患者的匿名数据集,包含27项基线特征及最佳药物治疗/戒烟依从性,比较机器学习、逻辑回归与8名临床医生的预测性能。
主要发现
- 机器学习在2年CLTI预测的AUC为0.885,逻辑回归为0.74,最佳临床医生为0.63;5年CLTI预测的AUC分别为0.936、0.808和0.639。
- 5年主要不良心血管事件预测的AUC:机器学习0.963,逻辑回归0.759,最佳临床医生0.611;主要不良肢体事件预测的AUC:机器学习0.957,逻辑回归0.9,最佳临床医生0.677。
- 机器学习模型Brier分数为0.03-0.07,逻辑回归为0.10-0.22,临床医生大于0.31;F1分数机器学习为0.80-0.86,逻辑回归为0.61-0.72,临床医生为0.50-0.59。
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
BACKGROUND: Recent evidence has shown that machine learning (ML) techniques can accurately forecast adverse cardiovascular and limb events in patients with intermittent claudication. This is the first study to compare the predictive performance of ML versus traditional logistic regression (LR) and clinicians. METHODS: An anonymized dataset of 99 patients with 27 baseline characteristics, compliance with best medical therapy/smoking cessation was used for comparison. Predictive performance was assessed using area under the receiver operating characteristic curve, F1 score, and Brier score. ML, LR, and clinicians were compared in their ability to predict outcomes including progression to chronic limb-threatening ischemia (CLTI) at 2 and 5 years, and probability of major adverse cardiovascular events or limb events upto 5 years. Independent variable importance ranking was performed to identify the most influential predictors. RESULTS: The Least Absolute Shrinkage and Selection Operator based ML model was compared with (LR) and predictions from 8 clinicians. ML significantly outperformed LR and clinicians across all outcomes. Area under the receiver operating characteristic curve for CLTI at 2 years: ML 0.885, LR 0.74, best clinician 0.63; CLTI at 5 years: ML 0.936, LR 0.808, best clinician 0.639; major adverse cardiovascular event at 5 years: ML 0.963, LR 0.759, best clinician 0.611; major adverse limb event: ML 0.957, LR 0.9, best clinician 0.677. Brier scores for the ML model demonstrated excellent accuracy: ML (0.03-0.07), compared to LR (0.10-0.22) and clinicians (>0.31).The ML model demonstrated superior predictive performance with F1 scores ranging from 0.80 to 0.86 across all outcomes, consistently outperforming both LR (F1 scores: 0.61-0.72) and individual clinicians (F1 scores: 0.50-0.59). CONCLUSION: ML-based prediction models significantly outperform traditional regression and clinician judgment, primarily due to their ability to capture complex nonlinear associations between variables.
引用
- Ravindhran B, Lim A, Pymer S, et al. Comparative Performance of Clinician and Computational Approaches in Forecasting Adverse Outcomes in Intermittent Claudication. Ann Vasc Surg. 2025. PMID: 40379095. doi:10.1016/j.avsg.2025.05.009
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。