MI-STEMI
Using Machine-learning and Image-recognition to improve automated ECG interpretation for patients with ST- Elevation Myocardial Infraction
INVESTIGATORS
Principal Investigator:
Dr. Sameer Masood, MD MPH FRCPC
Attending Emergency Physician
University Health Network
Assistant Professor, Department of Medicine, University of Toronto
sameer.masood@uhn.ca
Dr. Bo Wang PhD
Faculty Member, Vector Institute
AI Lead at PMCC, UHN
Scientist, Techna Institute, UHN
Assistant Professor, University of Toronto
Background/
Rationale
Obtaining and interpreting an ECG in a timely fashion is a vital part of the assessment of patients that present to the emergency department with cardiac complaints (Chest Pain, Shortness of breath, Palpitations, Dizziness, Syncope). Emergency Providers are expected to interpret an ECG within minutes and make critical treatment decisions for patients based on their interpretation of the ECG (1) . An important consideration for emergency providers is to rapidly diagnose patients with ST-Elevation Myocardial Infarction (STEMI), since timely diagnosis and treatment is associated with reduced morbidity and mortality (1) . Providers often rely on machine generated automated ECG interpretations to augment their clinical decision making. Automated ECGs were initially introduced to improve and standardize the diagnosis of ECGs by using pre-specified algorithms (2) . While automated ECG interpretation has been around since the 1950s, it’s uptake and utility has been limited due to lower accuracy compared to human interpretation (2) (3) . Several studies have shown that automated ECGs have significantly lower accuracy when diagnosing non-sinus rhythms (3) . Similarly, in diagnosing acute coronary syndromes (ACS), automated ECG interpretation algorithms show significant variation both with regards to false positives and false negatives, making their clinical utility limited (4) .
Machine learning is a group of computation technologies that allow for the recognition of patterns in data without explicitly having them programmed (supervised learning). These techniques can include logistic regression, support vector machines, and random-forests. Deep learning is an advancement of machine learning that uses programmatic representations of neural networks to process complex data. Recent work published in Nature showed that machine learning can achieve cardiologist level accuracy when detecting dysrhythmias through a single lead ECG (5) . Similarly, when applied to patients with ACS, machine learning has showed significant improvements over conventional automated ECG algorithms. However, previous work has largely focused on applying machine learning directly to digitized ECG signals and this presents inherent limitations in terms of their applicability to real-world scenarios as this requires buy in from various manufactures for implementation. An alternative to using direct ECG signals is to use image recognition. Using machine learning for image-recognition and classification has shown promise in dermatology (6). Applying it to ECG interpretation would allow us to develop a universal application that be used in all settings with access to a smartphone.
Automated Prognostication of safe discharge using Deep Learning Applied to Chest X-Rays of Patients with Suspected Pneumonia Presenting to the Emergency Department: A Shadow Deployment Study