Development and validation of warning system of ventricular tachyarrhythmia in patients with heart failure with heart rate variability data

Implantable-cardioverter defibrillators (ICD) detect and terminate life-threatening ventricular tachyarrhythmia with electric shocks after they occur. This puts patients at risk if they are driving or in a situation where they can fall. ICD’s shocks are also very painful and affect a patient’s quality of life. It would be ideal if ICDs can accurately predict the occurrence of ventricular tachyarrhythmia and then issue a warning or provide preventive therapy. Our study explores the use of ICD data to automatically predict ventricular arrhythmia using heart rate variability (HRV). A 5 minute and a 10 second warning system are both developed and compared. The participants for this study consist of 788 patients who were enrolled in the ICD arm of the Sudden Cardiac Death – Heart Failure Trial (SCD-HeFT). Two groups of patient rhythms, regular heart rhythms and pre-ventricular-tachyarrhythmic rhythms, are analyzed and different HRV features are extracted. Machine learning algorithms, including random forests (RF) and support vector machines (SVM), are trained on these features to classify the two groups of rhythms in a subset of the data comprising the training set. These algorithms are then used to classify rhythms in a separate test set. This performance is quantified by the area under the curve (AUC) of the ROC curve. Both RF and SVM methods achieve a mean AUC of 0.81 for 5-minute prediction and mean AUC of 0.87-0.88 for 10-second prediction; an AUC over 0.8 typically warrants further clinical investigation. Our work shows that moderate classification accuracy can be achieved to predict ventricular tachyarrhythmia with machine learning algorithms using HRV features from ICD data. These results provide a realistic view of the practical challenges facing implementation of machine learning algorithms to predict ventricular tachyarrhythmia using HRV data, motivating continued research on improved algorithms and additional features with higher predictive power.

植入式心律转复除颤器（Implantable-cardioverter defibrillators, ICD）可在致命性室性快速性心律失常发作后，通过电击检测并终止该类心律失常。但此举会使患者在驾驶或可能发生跌倒的场景中面临安全风险。此外，ICD放电带来的痛感极强，会严重影响患者的生活质量。若ICD能够精准预测室性快速性心律失常的发作，并提前发出预警或实施预防性治疗，将是理想的临床解决方案。本研究探索利用ICD数据，通过心率变异性（heart rate variability, HRV）自动预测室性心律失常。本研究开发并对比了两种预警系统：5分钟预警与10秒预警。本研究的参与者为788名纳入《心源性猝死-心力衰竭试验》（Sudden Cardiac Death – Heart Failure Trial, SCD-HeFT）ICD组的患者。研究分析了两类患者心律：正常窦性心律与室性快速性心律失常前驱心律，并提取了不同的心率变异性特征。研究在训练集子集的特征数据上训练了包括随机森林（random forests, RF）与支持向量机（support vector machines, SVM）在内的机器学习算法，以对两类心律进行分类。随后，将训练好的算法用于独立测试集的心律分类任务。模型性能通过受试者工作特征曲线（Receiver Operating Characteristic curve, ROC）下的面积（area under the curve, AUC）进行量化。随机森林与支持向量机模型在5分钟预测任务中均取得了0.81的平均曲线下面积；在10秒预测任务中，平均曲线下面积可达0.87~0.88。通常而言，曲线下面积超过0.8即具备进一步临床研究的价值。本研究表明，利用ICD数据中的心率变异性特征，通过机器学习算法即可实现中等精度的室性快速性心律失常预测。本研究结果为基于心率变异性数据、应用机器学习算法预测室性快速性心律失常的实际落地挑战提供了切实的认知，同时也为后续研发更高预测效能的优化算法与新增特征指明了研究方向。