Acceleration of machine learning research in healthcare is challenged by lack of large annotated datasets, and use of highly-biased corpora that rarely represent the entire data distribution. Consequently, there is a strong need for models that can generalize well to both unseen variations within the observed classes, and unseen classes. Furthermore, dealing with measurement inaccuracies and exploiting unsupervised data are considered to be central to designing next-generation solutions. In this work, we consider such a challenging problem in machine learning driven diagnosis – detecting a gamut of cardiovascular conditions (e.g. infarction, cardiomyopathy etc.) from limited channel ECG measurements. Lately, this problem has gained increased research interest with the wide-spread adoption of mobile patch-based ECG monitors to collect measurements from only a subset of channels. The recently concluded Physionet CinC challenge emphasized the importance of detecting cardiac diseases using single-channel ECG.
Though deep neural networks have achieved unprecedented success in predictive modeling, they rely solely on discriminative models that generalize poorly to unseen classes. We argue that unsupervised learning can be utilized to construct effective latent spaces that facilitate better generalization. We propose to employ unsupervised Seq2Seq models to construct latent spaces from limited channels, and subsequently use a state-of-the-art fully convolutional ResNet architecture for detection. This work extensively compares the generalization of our two-stage approach (ResNet++) against the ResNet architecture, to unseen disease conditions. Our results show a 2-11% improvement in F1-scores. For example, our results on the PTBDB dataset show that, ResNet++ trained to detect only inferior Myocardial Infarction (MI) using leads II, III, and aVF, is more effective than ResNet in detecting other MI variants (anterior, posterior, lateral), Bundle Branch Block, Dysrhythmia etc. Finally, through the use of unsupervised modeling, the proposed approach is less sensitive to hyper-parameters and produces models with low-variance in comparison to existing solutions.