Improving Automatic Clinical Decision Support System with Advanced Computational Methods

Abstract

Machine learning (ML) has become an essential component of clinical decision support systems, demonstrating significant effectiveness and potential across various tasks such as patient diagnosis, treatment recommendation, and patient outcome prediction. It effectively utilizes diverse data resources, including structured, unstructured, and also, multimodal data. The digitalization of medical data has facilitated enhanced resource prioritization and improved patient management. This integration of technology and healthcare has streamlined workflows and enabled more precise and personalized care strategies. As machine learning algorithms continue to evolve, they promise to further revolutionize healthcare by improving outcomes and optimizing the use of healthcare resources. These advancements are paving the way for more dynamic and responsive healthcare environments.

However, despite these technological developments, there remain challenges that need to be addressed to better harness the potential of machine learning in medicine. This dissertation investigates how to address three main technical challenges in developing and applying machine learning models to medical data. The first challenge is model generalizability, where the goal is to ensure that models perform consistently across different healthcare settings and patient populations, despite the variability in data sources and distributions. The second challenge is model interpretability, which is crucial for gaining trust from medical practitioners. Models should be able to provide explanations for their decisions in a way that aligns with medical knowledge and practice, facilitating their adoption in clinical decision-making. The third challenge involves the lack of data annotation, a common issue in medical datasets due to privacy concerns and the labor-intensive nature of obtaining detailed, accurate annotations from expert practitioners. Addressing these challenges is essential for the successful integration of machine learning technologies into practical and effective healthcare applications.

The structure of this dissertation is as follows: Chapter II introduces a novel regularized support vector machine that leverages partially available privileged information to enhance the model's generalizability. Chapter III explores the use of fuzzy neural networks to interpret data for predicting the necessity of advanced therapies in heart failure treatment. Chapter IV extends a novel algorithm introduced in Chapter II, which is capable of processing longitudinal data and offering interpretations at the rule level. Chapter V presents a framework that incorporates expert knowledge into large language models, achieving high accuracy while ensuring data privacy and improving latency. Furthermore, as a supplement, Chapter VI introduces (1) standard and personalized logistic tensor regression methods to improve model interpretability and (2) a three-stage framework that combines self-supervised learning with transfer learning to overcome medical imaging data scarcity. The final chapter summarizes the key findings and limitations of the studies conducted and outlines future research directions to further address these challenges.

In conclusion, this dissertation introduces algorithms and strategies designed to overcome existing technical challenges such as model generalizability, model interpretability, and the scarcity of data annotations in the integration of machine learning techniques with clinical support systems. The contributions of this work significantly enhance the application of ML-based systems in both clinical and research practices, paving the way for more robust, understandable, and efficient healthcare solutions.