A system approach to augment clinical decision-making using machine learning

Abstract

This thesis helps find limits within which automated methods can support and surpass the capabilities of medical professionals and the limits beyond which these methods are not yet adequate. This will inform later exploration about (a) what improvements in data collection, interpretation, and visualization will enhance technology's capacity and (b) what changes clinicians can make to improve their decision making-augmented or not. This thesis includes (a) describing clinical decisions, informed by literature and clinical case studies and (b) reviewing current capabilities of machine methods. This led to (c) a test experiment-how to use data about a particular condition (e.g. in-hospital mortality rate prediction) from a particular source (the MIMIC III data base). The results will help define current limits on augmenting clinical decisions and establish direction for future work including more demanding experiments.

Artificial Intelligence (AI) includes Machine Learning (ML), Natural Language Processing (NLP), Computer Vision, Speech Recognition, and Robotics. As an important branch of Al, ML builds statistical models to learn from sample data, known as "training", identifies patterns, and makes predictions based on new data, known as "inference." In this way, ML is useful in rationalizing and predicting in uncertain environments, with minimum human interventions. Decision making is central to the healthcare practice, with many decisions made under conditions of uncertainty. Clinicians must integrate a huge variety of data while pressured to decrease diagnostic uncertainties and risks to patients. Deciding what information to gather, which test to order, how to interpret and integrate this information to draw diagnostic conclusions, and which treatments to give are essential.

In typical situations, clinicians evaluate patient symptoms and potential disease patterns, confirmed by a variety of tests, and they initiate treatments based on their experience and customary practice. This is complicated when multiple illnesses coexist, the illness may be rare, the information may be conflicting, or prior interventions may affect the presenting symptoms.