Deep learning with limited labelled cardiac data

Cardiac data, that which pertain to the heart, are a rich source of information that reflect a patient's cardiac status. Extracting clinically-useful insight from such data can be achieved via deep learning, a sub-field of artificial intelligence. Current clinical deep learning algorithms, however, are heavily dependent upon resources such as abundant data and high-quality annotations, both of which are scarce in many clinical settings. In low-resource settings, for example, the prohibitively high cost of medical infrastructure precludes the collection of data and the limited number of physicians hampers the provision of annotations. The reasons for data paucity in high-resource settings are multi-fold, ranging from stringent patient-privacy regulations to the low inter-operability of medical records. Physicians in this setting can also be disengaged due to the overwhelming number of annotation requests. To address this challenge, in this thesis, we design deep learning algorithms that exploit cardiac data to achieve more with less; less data, fewer labels, and less medical supervision during the learning process. While designing these algorithms, we focus predominantly on the clinical task of cardiac arrhythmia classification, which involves diagnosing abnormalities in the functioning of the heart. We deploy our algorithms in three paradigms characterized by an incrementally increasing level of resource availability. In Part I of the thesis, we simulate a low-resource scenario with limited data and exploit conditional generative adversarial networks to generate cardiac time-series data for augmentation purposes. In Part II, we simulate access to abundant unlabelled data and limited labelled data. In this environment, we propose an active learning framework that dynamically determines whether an annotation should be requested from a physician or generated by an algorithm instead. We also present a family of patient-specific contrastive learning methods that improve resource-efficiency; the ability of a learner to solve a task with less data. In Part III, we deal with more extensive multi-modal data. We explore the degree to which algorithms suffer from catastrophic forgetting (the impaired ability to solve tasks from the past upon learning tasks in the present), and propose a continual learning framework to overcome this phenomenon. We also simultaneously exploit cardiac data and clinical textual reports to design a captioning system that, upon receiving cardiac signals, generates clinical reports in multiple languages. By designing such resource-efficient frameworks, we hope to improve the accessibility of clinical deep learning algorithms, and, in turn, healthcare to vulnerable patients in low-resource settings.