This is the first study that developed and validated a DLA for predicting cardiac arrest using ECGs. This study reveals that a deep-learning algorithm, one of the powerful tools of artificial intelligence, can figure out very delicate ECG changes in predicting cardiac arrest. The performance for predicting cardiac arrest was preserved with 6-lead or single-lead DLAs. In recent years, there many wearable devices for monitoring ECGs have been developed. If cardiac arrest could be predicted using ECGs, one can in principle capture the risk of patients in general wards or at home through wearable devices. Additionally, ECGs as raw bio-signal data can be used, which can enhance the performance of recent TTSs based on deep learning.
In development and validation using retrospective data of more than 46,000 ECGs, the DLA had a high AUROC of 0.913 to 0.948 for predicting cardiac arrest. The model showed good performance using data from another hospital that was not used for algorithm development, and had different patient characteristics and data shape. At a high-sensitivity (90%) operating point in development data, the DLA performed well as a potential TTS to predict cardiac arrest, and screen risk in patients with a negative predictive value greater than 99.8%. The model’s performance was better than conventional TTSs, and similar to recent novel TTSs based on deep learning. Also, the model’s performance was better than other commonly used screening tests such as mammography for breast cancer (AUROC, 0.78, positive predictive value, 3–12%), and fecal occult blood testing for detecting colorectal neoplasia (AUROC 0.71, overall sensitivity, 29%).[28, 29]
The most important aspect of deep learning is its ability to extract features and make an algorithm from various types of data, such as images, 2D data, and waveforms. Here, we used raw ECG data (2D numerical data, 12 × 4000) and interpreted ECG patterns for predicting cardiac arrest. Attia et al. developed deep-learning algorithms for screening cardiac contractile dysfunction, predicting the occurrence of atrial fibrillation during sinus rhythm, approximating age and sex, and detecting hyperkalemia using raw ECG data and demonstrated its feasibility.11,12 Our study group showed that a deep-learning-based algorithm using ECG could outperform cardiologists in diagnosing left ventricular hypertrophy and diagnosis aortic stenosis.13 However, deep learning is often criticized for the unreliability of its outcomes because of the unpredictability of the process. Because of this, we used a sensitivity map to visualize the regions of the ECGs that were used for decision-making by the DLA.
The map shows that the DLA focused more on the QRS complex to decide and predict cardiac arrest. The DLA also partially focused on the T-wave for predicting cardiac arrest. QRS prolongation has been considered a prognostic marker for mortality among patients with a variety of cardiovascular diseases.[16–18] Moreover, QRS fragmentation has been reported to be associated with increased mortality in patients with structural heart disease. In this study, the cardiac arrest ECGs had prolonged QRS durations and QTs corrected. The heart rates of cardiac arrest ECGs were greater than that of nonevent ECGs, and the T wave axes of cardiac arrest ECGs were more rightward than that of nonevent ECGs.
Our study has several limitations to be resolved in the future. First, this was a retrospective study using conventional 12-lead ECGs. A prospective study is warranted to determine the association of the DLA, and enhancement in detecting cardiac arrest and improving clinical outcomes. A study for confirming the accuracy of data from various wearable or portable ECG devices is warranted to apply the DLA to those devices. If we adopt DLAs in daily living, a study is also needed to confirm the performance at home and general environments. Second, the performance of the DLA needs to be enhanced in order to use it as a reliable cardiac arrest detecting tool. Although the NPV was over 99%, the PPV was only 8% at the point of high sensitivity. As there are several methodologies that have been developed in deep learning and computer science, we could develop higher performance DLAs in the near future. We also developed a high performance TTS, by combining the DLA with discrete numeric variables such as the vital sign. Thirdly, we need to explore the decision-making process of the DLA further. For example, additional experiments are required to understand the deep learning process better, and thereby understand which exact characteristics of the QRS complex and the T wave influence the algorithm’s decision. Explainable artificial intelligence has been studied and reported on recently, so the “black box” limitation could be solved in the near future. This subject will be our next area of study, and this might turn out to be the new standard for discovering new medical knowledge about diseases and ECGs.