Machine learning-based prediction of heart failure readmission or death: implications of choosing the right model and the right metrics
Machine learning-based prediction of heart failure readmission or death: implications of choosing the right model and the right metrics
复制标题
DOI：
10.1002/ehf2.12419
复制
发表时间：
2019-04-01
影响因子：
3.8
通讯作者：
Dwivedi, Girish
中科院分区：
文献类型：
--
作者：
Awan, Saqib Ejaz;Bennamoun, Mohammed;Dwivedi, Girish
AIMS机器学习(ML)被广泛认为能够从数据中学习复杂的隐藏交互作用，并具有预测心力衰竭(HF)、再入院和死亡等事件的潜力。最近的研究显示，相互矛盾的结果可能是因为没有考虑到医学数据中常见的阶层失衡问题。我们开发了一种新的ML方法来预测30天再次入院或死亡，并将该模型的性能与其他常用的预测模型进行了比较。方法和结果我们在链接的医院发病率数据收集中确定了2003至2008年间所有65岁以上的西澳大利亚患者入院的心衰患者。考虑到类别失衡的问题，我们开发了一种基于多层感知器(MLP)的方法来预测30天的HF再入院或死亡，并使用性能指标比较了预测性能指标，即接收器操作特征曲线下的面积(AUC)、精确召回曲线下的面积(AUPRC)、灵敏度和特异度与其他ML和回归模型。在10757名心力衰竭患者中，23.6%的患者在出院后30天内再次入院或死亡。我们观察到加权随机森林、加权决策树、Logistic回归和加权支持向量机模型的AUC分别为0.55、0.53、0.58和0.54，而AUPRC分别为0.39、0.38、0.46和0.38。基于MLP方法的AUC(0.62)和AUPRC(0.46)最高，其敏感度为48%，特异度为70%。结论对于有类别不平衡的医疗数据，基于MLP的方法在预测30天心力衰竭再入院或死亡方面优于其他的ML和回归方法。
Aims Machine learning (ML) is widely believed to be able to learn complex hidden interactions from the data and has the potential in predicting events such as heart failure (HF) readmission and death. Recent studies have revealed conflicting results likely due to failure to take into account the class imbalance problem commonly seen with medical data. We developed a new ML approach to predict 30 day HF readmission or death and compared the performance of this model with other commonly used prediction models.Methods and results We identified all Western Australian patients aged above 65 years admitted for HF between 2003 and 2008 in the linked Hospital Morbidity Data Collection. Taking into consideration the class imbalance problem, we developed a multi-layer perceptron (MLP)-based approach to predict 30 day HF readmission or death and compared the predictive performances using the performance metrics, that is, area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), sensitivity and specificity with other ML and regression models. Out of the 10 757 patients with HF, 23.6% were readmitted or died within 30 days of hospital discharge. We observed an AUC of 0.55, 0.53, 0.58, and 0.54 while an AUPRC of 0.39, 0.38, 0.46, and 0.38 for weighted random forest, weighted decision trees, logistic regression, and weighted support vector machines models, respectively. The MLP-based approach produced the highest AUC (0.62) and AUPRC (0.46) with 48% sensitivity and 70% specificity.Conclusions We show that for the medical data with class imbalance, the proposed MLP-based approach is superior to other ML and regression techniques for the prediction of 30 day HF readmission or death.