Predicting adverse cardiac events at the emergency department: A deep learning approach

The emergency department is a stressful environment, in which physicians are required to make fast and accurate diagnostic assessments amidst an ever increasing flood of clinical information, including a growing body of medical knowledge. Meanwhile, the digitization of medical health records in combination with recent breakthroughs in Artificial Intelligence is ushering in a new era of precision medicine. Deep-learning powered decision support tools represent a promising avenue for improving patient outcomes and reducing work-flow complexity for physicians. The goal of this thesis was to explore ways to apply modern deep-learning algorithms to improve predictions of adverse cardiac events among chest-pain patients at the emergency department.

The first paper investigated the utility of prior electrocardiograms (ECGs) for predicting major adverse cardiac events (MACE). We found that, contrary to clinical recommendations, prior ECGs will not meaningfully contribute to the predictions of MACE. The second paper aimed at quantifying the benefit of transfer learning for models using ECGs to predict acute myocardial infarction (AMI). We found that a simple transfer-learning strategy of pre-training on patient age and sex resulted in a substantial increase in the downstream performance of AMI, while simultaneously enabling the use of larger and more powerful deep-learning model architectures. The third paper explored options for early rule-out of AMI. The results indicated that as many as 16% of chest-pain patients can be safely ruled out based on age and sex alone, suggesting opportunities to further streamline the clinical pathway for low-risk patients. The fourth paper was concerned with identifying and predicting acute coronary occlusion myocardial infarctions (OMI), which are particularly serious and require urgent invasive treatment. We found that 29% of all AMI patients could be classified as OMI, but only 11% of those patients received timely treatment. Our deep-learning ECG model was able to predict the OMI outcome with an AUC of 88% using only the ECG and medical history, and 95.3% when including the initial high-sensitivity cardiac troponin T lab results.

In conclusion, this thesis identified several promising applications and improvements of deep-learning algorithms in predicting adverse cardiac events in chest-pain patients. Much work remains to be done, including prospective clinical trials, but hopefully we are now one step closer to the implementation of a deep-learning powered decision support tool that can help save lives at the emergency department.