Principled Deep Learning for Healthcare Applications

Abstract

Healthcare stands to benefit from the advent of deep learning on account of (i) the massive amounts of data generated by the health system and (ii) the ability of deep models to make predictions from complex inputs. This dissertation centers on two applications of deep learning to challenging problems in healthcare.First, we discuss deep learning for treatment effect/counterfactual estimation in the observational setting, i.e., where the treatment assignment is not randomized (Chapters 2 and 3). For example, we may want to know the causal effect of a drug on a patient's blood pressure. We combine deep learning with classical weighting techniques to estimate average and conditional average treatment effects from observational data. We show theoretical properties of our method, including guarantees about when "balance" can be achieved between treatment groups. We then weaken the typical "ignorability" assumption and generate treatment effect intervals (instead of point-estimates).Second, we explore the use of deep learning applied to a difficult problem in medical imaging: classifying malignancy from thyroid cytopathology slides (Chapters 4, 5, and 6). The difficulty of this problem arises from the image size, which is typically on the order of tens of gigabytes (i.e., around 3 to 4 orders of magnitude larger than image sizes in popular deep learning architectures). Our approach is a two-step process: (i) automatically finding image regions containing follicular cell groups, (ii) classifying each region and aggregating the predictions. We show that our system works well for mobile phone images of thyroid biopsy slides, and that our system compares favorably with state-of-the-art genetic testing for malignancy.Finally, after my Ph.D. I plan to enter a career in autonomous driving. As an "epilogue" of this dissertation (Chapter 7), we present a method to make deep learning point-cloud models for autonomous driving which are invariant (or equivariant) to rotations. Intuitively, this is an important requirement -- a rotated bicycle should still be classified as a bicycle, and driving behavior should be independent of direction of travel. However, most deep learning models used in autonomous driving today do not satisfy these properties exactly. We propose a practical model (based on the Transformer architecture) to address this pitfall, and we showcase its performance on point-cloud classification and trajectory forecasting tasks.