Deep learning on ECGs to improve outcomes in patients on dialysis

摘要。 透析中低血压（IDH）和主要不良心血管事件（MACE）在接受 维持性血液透析（HD），并显著增加这一脆弱患者的发病率和死亡率 人口虽然存在减少这些不良后果的策略，但缺乏准确和可操作的方法， 预测性风险模型导致这些战略的总体利用率较低且无针对性。 心电图（ECG）是普遍存在的、便宜的、执行简单的，并且它提供了一种立即可获得的， 非侵入性地洞察心血管反射和健康。原始波形数据可以通过以下方式利用 先进的深度学习，用于准确确定各种心脏特征， 关键成果。在我们之前发表的工作中，我们展示了深度学习的实用性，以确定两者的正确性。 和左心功能以及迁移学习在改善HD患者预后预测中的作用。在 最近的初步分析，我们还显示了实用的波形数据，以预测在医院IDH和协会 30天死亡率。然而，IDH的未来发展和验证， MACE对于临床部署至关重要。因此，扩展我们以前的工作，我们提出了最大的前景 利用ECG预测HD患者关键结局的研究。我们将招募1000名不同的患者 来自纽约市透析单位的HD患者（推导）和来自北卡罗来纳州的150例患者（验证）， 在基线和基线后4周获得标准持续时间、12导联ECG。此外， 参与者将在3个连续HD疗程期间接受连续波形监测， 子研究。然后，我们将使用深度学习和迁移学习（使用我们的预训练模型）。 大约1100万个档案ECG数据库），并使用该数据预测同一会话中的IDH，并在30 天（目标1）和1年随访时MACE的复合结局（目标2）。这一提议的结果 对于短期和长期心脏结局的预测具有很高的临床重要性。积极成果 将推动研究测试部署我们的预测模型到HD单位，以检测和预防 IDH和MACE以及使用新型可穿戴设备进行IDH和心脏风险预测。

ABSTRACT. Intradialytic hypotension (IDH) and major adverse cardiovascular events (MACE) are common in patients on maintenance hemodialysis (HD) and contribute significantly to morbidity and mortality in this vulnerable patient population. Although strategies to decrease these adverse outcomes exist, the lack of accurate and actionable predictive risk models has led to overall low and non-targeted utilization of these strategies. Electrocardiography (ECG) is ubiquitous, cheap, simple to perform, and it provides an immediately accessible, non-invasive insight into cardiovascular reflexes and health. The raw waveform data can be leveraged by advanced deep learning for accurate determination of various cardiac features as well as prognostication of key outcomes. In our prior published work, we demonstrated the utility of deep learning to determine both right and left heart function and the utility of transfer learning to improve outcome prediction in patients on HD. In recent preliminary analysis, we also show utility of waveform data to predict in hospital IDH and association with 30-day mortality using retrospective data. However, prospective development and validation on IDH and MACE are critical to clinical deployment. Thus, extending our prior work, we propose the largest prospective study on utilizing ECGs for prediction of key outcomes in patients on HD. We will recruit 1000 diverse patients on HD from dialysis units in New York City (derivation) and 150 patients from North Carolina (validation) and obtain standard duration, 12-lead ECGs at baseline and 4 weeks after baseline. In addition, a subset of participants will undergo continuous waveform monitoring during 3 consecutive HD sessions in an exploratory sub-study. We will then use deep learning and transfer learning (using pre-trained models from our approximately 11 million archival ECG database) and use this to predict IDH at the same session and within 30 days (Aim 1) and a composite outcome of MACE at 1 year of follow up (Aim 2). The results of this proposal are of high clinical importance for the prediction of both short- and long-term cardiac outcomes. Positive results will prompt studies testing deployment of our predictive models into HD units for detection and prevention of IDH and MACE as well use of novel wearables for IDH and cardiac risk prediction.