| Summary: | Machine learning models in healthcare have been widely studied in a number of contexts ranging from clinical risk stratification to image-guided diagnosis and prognostication. Nevertheless, key challenges remain from both clinical and technical perspectives. In the case of prediction models, for example, predicting the occurrence of rare clinical events is often challenging, mainly because of extreme class imbalance in the training data. Estimating treatment effect, on the other hand, is hindered by the fact that the common support assumption is not \textit{a priori} guaranteed to be valid in non-randomized data. This thesis develops and applies approaches that address these challenges in order to obtain clinically useful insights.
In the first part of the thesis, we tackle these obstacles in the context of Acute Coronary Syndrome (ACS) - a condition where blood flow to the heart suddenly becomes compromised. We use a contrastive Variational Autoencoder (contrastive-VAE), an approach that models both the majority and minority classes as having shared latent properties, to address the following challenges: 1) Predicting rare adverse clinical outcomes after ACS; 2) Quantifying common support for estimating the effect of therapies for ACS; and 3) Causal feature selection for estimating individual treatment effects (ITE). For the first challenge, we demonstrate that generative oversampling with a contrastive-VAE significantly improves the discriminatory ability of predictive models relative to other traditional methods like SMOTE (Synthetic Minority Oversampling Technique). Similarly, for the problem of common support estimation, we show that a contrastive-VAE can effectively model the overlap between multiple treatment groups, yielding a quantitative estimate of the common support for the individual treatment effect and concomitant confidence intervals for the ITE estimate. Lastly, by modeling the joint distribution of patient features, treatments, and outcomes, we demonstrate that one can effectively identify a subset of patient features that are most important for ITE estimation, and that this smaller subset yields more precise ITEs with smaller confidence intervals.
In the second part of the thesis, we turn to a challenging clinical problem that uses ultrasound imaging for diagnosis and prognostication. Cardiac ultrasound (or echocardiography) plays a central role in the diagnosis and management of patients with suspected aortic stenosis (AS) - a disorder where one of the valves in the heart does not fully open. A complete echocardiographic study is typically performed by a trained sonographer who acquires videos of multiple views of the heart, and echocardiographers (cardiologists who specialize in the analysis of echocardiograms) interpret these videos, yielding clinically useful information. To facilitate the acquisition and interpretation of echocardiographic data, we developed a deep learning model that uses a single echocardiographic view (as opposed to use all of the acquired views) to diagnoses severe AS. We trained and evaluated the model based on spatial-temporal convolution that can accurately identify two key indicators of severe AS: large mean gradient over valve (0.88 AUC) and narrowed aortic valve area (0.78 AUC). Our approach might enable early detection of severe AS by non-specialists.
|
|---|