Cardiac MRI allows for a comprehensive assessment of myocardial structure, function, and tissue characteristics. Here we describe a foundational vision system for cardiac MRI, capable of representing the breadth of human cardiovascular disease and health. Our deep learning model is trained via self-supervised contrastive learning, by which visual concepts in cine-sequence cardiac MRI scans are learned from the raw text of the accompanying radiology reports. We train and evaluate our model on data from four large academic clinical institutions in the United States. We additionally showcase the performance of our models on the UK BioBank, and two additional publicly available external datasets. We explore emergent zero-shot capabilities of our system, and demonstrate remarkable performance across a range of tasks; including the problem of left ventricular ejection fraction regression, and the diagnosis of 35 different conditions such as cardiac amyloidosis and hypertrophic cardiomyopathy. We show that our deep learning system is capable of not only understanding the staggering complexity of human cardiovascular disease, but can be directed towards clinical problems of interest yielding impressive, clinical grade diagnostic accuracy with a fraction of the training data typically required for such tasks.
Wearable devices enable theoretically continuous, longitudinal monitoring of physiological measurements like step count, energy expenditure, and heart rate. Although the classification of abnormal cardiac rhythms such as atrial fibrillation from wearable devices has great potential, commercial algorithms remain proprietary and tend to focus on heart rate variability derived from green spectrum LED sensors placed on the wrist where noise remains an unsolved problem. Here, we develop a multi-task deep learning method to assess signal quality and arrhythmia event detection in wearable photoplethysmography devices for real-time detection of atrial fibrillation (AF). We train our algorithm on over one million simulated unlabeled physiological signals and fine-tune on a curated dataset of over 500K labeled signals from over 100 individuals from 3 different wearable devices. We demonstrate that in comparison with a traditional random forest-based approach (precision:0.24, recall:0.58, f1:0.34, auPRC:0.44) and a single task CNN (precision:0.59, recall:0.69, f1:0.64, auPRC:0.68) our architecture using unsupervised transfer learning through convolutional denoising autoencoders dramatically improves the performance of AF detection in participants at rest (pr:0.94, rc:0.98, f1:0.96, auPRC:0.96). In addition, we validate algorithm performance on a prospectively derived replication cohort of ambulatory subjects using data derived from an independently engineered device. We show that two-stage training can help address the unbalanced data problem common to biomedical applications where large well-annotated datasets are scarce. In conclusion, though a combination of simulation and transfer learning and we develop and apply a multitask architecture to the problem of AF detection from wearable wrist sensors demonstrating high levels of accuracy and a solution for the vexing challenge of mechanical noise.