Deep learning and hyperparameter optimization for assessing one's eligibility for a subcutaneous implantable cardioverter-defibrillator
Deep learning and hyperparameter optimization for assessing one's eligibility for a subcutaneous implantable cardioverter-defibrillator
复制标题
DOI：
10.1007/s10479-023-05326-1
复制
发表时间：
2023-05-24
影响因子：
4.8
通讯作者：
Zemkoho，Alain B.
中科院分区：
文献类型：
--
作者：
Dunn，Anthony J.;Coniglio，Stefano;Zemkoho，Alain B.
对于属于高危人群的室性心律失常（心源性猝死的主要原因）患者，通过植入皮下植入式心律转复除颤器（S-ICDs）进行治疗是标准的心脏病学实践。s - icd具有所谓的T波感应（TWOS）风险，这可能导致不适当的冲击，从而带来固有的健康风险。出于这个原因，根据目前的实践，患者的心电图（ECGs）是由心脏病专家在10秒内手工筛选的，以评估T:R比——T波和R波振幅之间的比率，用于作为twos可能性的标记——使用塑料模板。不幸的是，患者T:R比的时间变异性可能导致这种筛查程序不可靠，这种筛查程序由于其人工性质而不可避免地依赖于短的ECG段。在本文中，我们提出并研究了一种基于深度学习的工具，用于自动预测能够进行24小时自动筛选的多个10秒ECG记录片段的T:R比率。由于大大增加了筛选窗口，这种筛选将提供比目前使用的基于模板的10秒手动筛选更可靠的T:R比率预测。据我们所知，我们的工具是第一个完全自动化这种手动且可能不准确的过程的工具。从方法学的角度来看，我们为我们的工具评估了不同的深度学习模型架构，评估了一系列基于随机梯度下降的优化方法，用于训练其底层深度学习模型，执行超参数调优，并创建了性能最佳的模型的集合，以确定哪种组合能带来最佳性能。我们发现，所得到的模型已被集成到临床医生使用的原型工具中，能够非常准确地预测T:R比率。正因为如此，我们的自动化T:R比检测工具将使临床医生能够提供一个完全自动化的评估病人是否有资格植入S-ICD，这比目前的做法更可靠，因为采用了一个明显更长的心电图筛查窗口，比目前的人工做法更好、更准确地捕捉病人的T:R比的行为。
It is standard cardiology practice for patients suffering from ventricular arrhythmias (the main cause of sudden cardiac death) belonging to high risk populations to be treated via the implantation of Subcutaneous Implantable cardioverter-defibrillators (S-ICDs). S-ICDs carry a risk of so-called T wave over sensing (TWOS), which can lead to inappropriate shocks that carry an inherent health risk. For this reason, according to current practice patients’ Electrocardiograms (ECGs) are manually screened by a cardiologist over 10 s to assess the T:R ratio—the ratio between the amplitudes of the T and R waves which is used as a marker for the likelihood of TWOS—with a plastic template. Unfortunately, the temporal variability of a patient’ T:R ratio can render such a screening procedure, which relies on an inevitably short ECG segment due to its manual nature, unreliable. In this paper, we propose and investigate a tool based on deep learning for the automatic prediction of the T:R ratios from multiple 10-second segments of ECG recordings capable of carrying out a 24-hour automated screening. Thanks to the significantly increased screening window, such a screening would provide far more reliable T:R ratio predictions than the currently utilized 10-second, template-based, manual screening is capable of. Our tool is the first, to the best of our knowledge, to fully automate such an otherwise manual and potentially inaccurate procedure. From a methodological perspective, we evaluate different deep learning model architectures for our tool, assess a range of stochastic-gradient-descent-based optimization methods for training their underlying deep-learning model, perform hyperparameter tuning, and create ensembles of the best performing models in order to identify which combination leads to the best performance. We find that the resulting model, which has been integrated into a prototypical tool for use by clinicians, is able to predict T:R ratios with very high accuracy. Thanks to this, our automated T:R ratio detection tool will enable clinicians to provide a completely automated assessment of whether a patient is eligible for S-ICD implantation which is more reliable than current practice thanks to adopting a significantly longer ECG screening window which better and more accurately captures the behavior of the patient’s T:R ratio than the current manual practice.