机器学习分层阻塞性睡眠呼吸暂停患者的心血管疾病风险
本研究基于多种族动脉粥样硬化研究(MESA)中967名45至84岁基线无OSA或CVD的成人,回顾性分析基线临床因素,构建六种机器学习模型预测OSA患者未来10年CVD风险。最佳模型随机森林和bootstrap聚合的平衡准确率分别为91%和92%,最强预测因子为空腹血糖>91 mg/dL、舒张压>73 mm Hg和年龄>59岁。
要点
- 空腹血糖>91 mg/dL、舒张压>73 mm Hg和年龄>59岁是OSA患者发生CVD的最强预测因子。
- 随机森林和bootstrap聚合模型预测OSA患者CVD风险的平衡准确率分别为91%和92%。
- 基于基线临床因素构建的机器学习模型有助于OSA患者的心血管风险分层和早期干预。
核心结论
- 在基线无OSA或CVD的成人中,OSA患者未来发生CVD的最强预测因子为空腹血糖、舒张压和年龄,机器学习模型可用于心血管风险分层。
研究设计与样本
- 回顾性分析前瞻性收集的数据,纳入967名45至84岁参与多种族动脉粥样硬化研究(MESA)的成人。
- 使用逐步变量选择初步确定的基线临床因素构建六种机器学习模型,预测OSA患者额外CVD风险。
- 采用局部可解释模型无关解释(LIME)评估模型可解释性。
主要发现
- 967名基线无OSA或CVD的成人中,入组10年后116名诊断为OSA和CVD,851名仅诊断为OSA。
- 最佳模型随机森林的敏感性84%、特异性99%、平衡准确率91%。
- bootstrap聚合模型的敏感性84%、特异性100%、平衡准确率92%。
- OSA合并CVD versus 仅OSA的最强预测因子为空腹血糖>91 mg/dL、舒张压>73 mm Hg和年龄>59岁。
临床怎么用
- 对于OSA患者,可关注空腹血糖>91 mg/dL、舒张压>73 mm Hg和年龄>59岁等基线因素以识别CVD高风险个体。
- 机器学习模型(如随机森林或bootstrap聚合)可能辅助OSA患者的心血管风险分层,指导早期干预。
局限性
- 本研究为回顾性分析,证据等级为3级。
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
OBJECTIVES/HYPOTHESIS: Obstructive sleep apnea (OSA) is associated with higher risk of morbidity and mortality related to cardiovascular disease (CVD). Due to overlapping clinical risk factors, identifying high-risk patients with OSA who are likely to develop CVD remains challenging. We aimed to identify baseline clinical factors associated with the future development of CVD in patients with OSA. STUDY DESIGN: Retrospective analysis of prospectively collected data. METHODS: We performed a retrospective analysis of 967 adults aged 45 to 84 years and enrolled in the Multi-Ethnic Study of Atherosclerosis. Six machine learning models were created using baseline clinical factors initially identified by stepwise variable selection. The performance of these models for the prediction of additional risk of CVD in OSA was calculated. Additionally, these models were evaluated for interpretability using locally interpretable model-agnostic explanations. RESULTS: Of the 967 adults without baseline OSA or CVD, 116 were diagnosed with OSA and CVD and 851 with OSA alone 10 years after enrollment. The best performing models included random forest (sensitivity 84%, specificity 99%, balanced accuracy 91%) and bootstrap aggregation (sensitivity 84%, specificity 100%, balanced accuracy 92%). The strongest predictors of OSA and CVD versus OSA alone were fasting glucose >91 mg/dL, diastolic pressure >73 mm Hg, and age >59 years. CONCLUSION: In the selected study population of adults without OSA or CVD at baseline, the strongest predictors of CVD in patients with OSA include fasting glucose, diastolic pressure, and age. These results may shape a strategy for cardiovascular risk stratification in patients with OSA and early intervention to mitigate CVD-related morbidity. LEVEL OF EVIDENCE: 3 Laryngoscope, 132:234-241, 2022.
引用
- Gourishetti SC, Taylor R, Isaiah A, et al. Stratifying the Risk of Cardiovascular Disease in Obstructive Sleep Apnea Using Machine Learning. Laryngoscope. 2021. PMID: 34487556. doi:10.1002/lary.29852
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。