Trustworthy Machine Learning for Clinical Diagnosis and Decision Support

ABSTRACT Infectious diseases, such as pulmonary tuberculosis and sepsis, are associated with significant patient morbidity and mortality. Because of their public health significance, recent research in artificial intelligence (AI) and machine learning (ML) have explored how deep learning models may be used as clinical decision support tools to work alongside clinicians in improving patient care. However, it is well-documented that AI algorithms—even those approved by the Food and Drug Administration (FDA)—are inaccurate and perform poorly on real-world patients, especially those that are not made sufficiently available during model development. As a result, such tools often inadvertently lack consistency and reliability due to their poor performance on different patient populations. Especially in high-stakes applications such as healthcare where there is a small margin for error, it is important to train neural networks that are clinically interpretable, trustworthy, and reliable for all patient populations. The reasons behind inconsistent model performance are complex, but can be largely distilled into two major algorithmic limitations: (1) traditional ML predictive models may not be well-calibrated with true clinical decision-making used by physicians; and (2) publicly available model training datasets may lack the natural variation in patient disease presentation necessary to ensure generalizable model performance. In this work, I propose a series of algorithmic and practical innovations to address these two algorithmic limitations in the clinical diagnosis and management of pulmonary tuberculosis and sepsis. I will accomplish this task by increasing the robustness and reliability of AI models in healthcare settings. Firstly, in order to better align predictive models with clinical reasoning for the diagnosis of pulmonary tuberculosis, I will show how publicly available vision-language foundational models can be used to improve the diagnostic accuracy of clinicians in resource-limited hospital settings (Aim 1). Secondly, I will propose and validate a novel computational algorithm that leverages historical aggregate patient data to improve the reliability of clinical decision support systems for sepsis treatment (Aim 2). These experiments will collectively demonstrate how AI algorithms can be better leveraged for various applications spanning multiple domains of patient care. In providing solutions for these clinically relevant problems, I will further develop the technical skills and scientific reasoning needed as a future radiologist and academic researcher in machine learning. I look forward to leveraging both my graduate education and clinical training together to ultimately become a well-rounded physician-scientist and independent investigator.