Date of Award
12-22-2023
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
In the era of digital health, Electronic Health Records (EHR) integrate clinical data from various sources, presenting a wealth of information across multiple modalities. While this abundance allows for comprehensive patient characterization, the intricate nature and high dimensionality of EHR data pose challenges for clinicians, particularly in the dynamic landscape of clinical decision-making. This dissertation delves into the application of Artificial Intelligence (AI) and Machine Learning (ML) to automate risk identification and provide decision support in clinical settings. The research unfolds in three key contributions. First, we investigate ML's potential in constructing a multi-modal framework for generating informative risk assessments. This framework incorporates specialized ML components to capture data representations from diverse modalities such as clinical texts, time series, and static features (numerical or categorical). Identifying gaps in existing models, we propose a semi-supervised, disentangled generative modelling approach to unveil nonlinear relationships within the high-dimensional and partially observable inputs, facilitating enhanced predictive capabilities. This work has been applied to and evaluated in clinical studies for predicting postoperative complications in perioperative care. Second, we extend ML techniques beyond predicting patient outcomes to predicting individualized treatment effects. This supports clinician decisions on personalized treatment for individual patients. Recognizing the need for both factual and counterfactual risk predictions under the constraints of input complexity and data imbalance, we introduce a semi-supervised generative representation framework for estimating individualized treatment effects. This work has been used to predict the need for critical life support for COVID-19 patients and validated on both an international dataset and the dataset of a local healthcare system. Finally, we capitalize on recent advancements in Large Language Modeling (LLM) as we revisit the learning of clinical texts. Validating the potential of pretrained LLMs in perioperative care, we identify an optimal transfer learning strategy for clinical LLMs to augment clinical decision-making processes. The effectiveness of this approach has been demonstrated on a large perioperative dataset. This research underscores the transformative role of AI and ML in harnessing the potential of vast EHR datasets, offering innovative solutions for risk identification and informed clinical decision-making in various healthcare applications.
Language
English (en)
Chair
Chenyang Lu