Shih Ling-Chieh, Wang Yu-Ching, Hung Ming-Hui, Cheng Han, Shiao Yu-Chieh, Tseng Yu-Hsuan, Huang Chin-Chou, Lin Shing-Jong, Chen Jaw-Wen
School of Medicine, College of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan.
Department of Medical Education and Research, Kaohsiung Veterans General Hospital, Kaohsiung, Taiwan.
Eur Heart J Digit Health. 2022 Nov 8;3(4):559-569. doi: 10.1093/ehjdh/ztac066. eCollection 2022 Dec.
The detection of white-coat hypertension/white-coat uncontrolled hypertension (WCH/WUCH) with out-of-office blood pressure (BP) monitoring is time- and resource-consuming. We aim to develop a machine learning (ML)-derived prediction model based on the characteristics of patients from a single outpatient visit.
Data from two cohorts in Taiwan were used. Cohort one (970 patients) was used for development and internal validation, and cohort two (464 patients) was used for external validation. WCH/WUCH was defined as an office BP of ≥140/90 mmHg and daytime ambulatory BP of <135/85 mmHg in treatment-naïve or treated individuals. Logistic regression, random forest (RF), eXtreme Gradient Boosting, and artificial neural network models were trained using 26 patient parameters. We used SHapley Additive exPlanations values to provide explanations for the risk factors. All models achieved great area under the receiver operating characteristic curve (AUROC), specificity, and negative predictive value in both validations (AUROC = 0.754-0.891; specificity = 0.682-0.910; negative predictive value = 0.831-0.968). The RF model was the best performing (AUROC = 0.884; sensitivity = 0.619; specificity = 0.887; negative predictive value = 0.872; accuracy = 0.819). The five most influential features of the RF model were office diastolic BP, office systolic BP, current smoker, estimated glomerular filtration rate, and fasting glucose level.
Our prediction models achieved good performance, underlining the feasibility of applying ML models to outpatient populations for the diagnosis of WCH and WUCH. Further validation with other prospective data sets should be considered in the future.
采用诊室外血压监测来检测白大衣高血压/白大衣未控制高血压（WCH/WUCH）既耗费时间又需要资源。我们旨在基于单次门诊患者的特征开发一种机器学习（ML）衍生的预测模型。
使用了台湾两个队列的数据。队列一（970例患者）用于模型开发和内部验证，队列二（464例患者）用于外部验证。WCH/WUCH定义为初治或已治疗个体的诊室血压≥140/90 mmHg且日间动态血压<135/85 mmHg。使用26个患者参数训练逻辑回归、随机森林（RF）、极端梯度提升和人工神经网络模型。我们使用夏普利值（SHapley Additive exPlanations values）来解释危险因素。在两次验证中，所有模型在受试者工作特征曲线下面积（AUROC）、特异性和阴性预测值方面均表现出色（AUROC = 0.754 - 0.891；特异性 = 0.682 - 0.910；阴性预测值 = 0.831 - 0.968）。RF模型表现最佳（AUROC = 0.884；灵敏度 = 0.619；特异性 = 0.887；阴性预测值 = 0.872；准确率 = 0.819）。RF模型最具影响力的五个特征是诊室舒张压、诊室收缩压、当前吸烟者、估计肾小球滤过率和空腹血糖水平。
我们的预测模型性能良好，突显了将ML模型应用于门诊人群诊断WCH和WUCH的可行性。未来应考虑使用其他前瞻性数据集进行进一步验证。
