Machine Learning and Image Processing for Clinical Outcome Prediction: Applications in Medical Data from Patients with Traumatic Brain Injury, Ulcerative Colitis, and Heart Failure

Abstract

Artificial intelligence (AI) and machine learning (ML) have achieved extensive success in many fields. They are powerful in pattern recognition and function modeling. The digitization of health data provides an important opportunity for improving care delivery and patient management through the AI-based clinical decision-support (CDS) system. Medical images are important components in evaluating the disease severity. While the human’s interpretation of medical images is subjective and qualitative, AI-based models can analyze those data in a more reproducible, quantitative, and less expensive way. With clinical observations and quantitative findings extracted from medical images, ML methods can be used to learn and discover knowledge. The automated CDS system can provide recommendations on diagnosis, treatment, and outcome prediction by leveraging massive medical data. Those systems can facilitate drug development, disease pathology research, and clinical practice.

This dissertation investigates medical image analysis and CDS systems development in a more reliable, interpretable manner. Limitations exist in applying AI/ML techniques in medical problems. Medical data may have high variability in terms of the patient population, collection site, equipment, and imaging protocols. It is crucial that the ML and image processing algorithms have a good generalizability and can be reliably applied to unseen patient data. In addition, a broad spectrum of AI/ML methods is among the “black box” models. The lack of justification leads to concerns and hesitations of using AI/ML techniques in clinical or research practice. Features with clinical meaning and models that can be well explained can gain more trust and are more favorable to end-users. In this dissertation, several AI-based CDS systems have been designed and implemented to facilitate clinical and research practice. Novel algorithms are proposed to overcome the challenges of applying AI/ML techniques. To improve the generalizability of the deep learning models, a robust learning algorithm is proposed to encourage the network to be invariant to hematoma intensity variability. A Scale Module and filter pruning technique are proposed to reduce the network’s size and complexity. To improve the interpretability of the CDS systems, a transparent ML algorithm is proposed based on tropical geometry and fuzzy logic, which can learn humanly understandable rules from the dataset and integrate existing domain knowledge to facilitate the model training. Domain knowledge plays an important role in the design of CDS systems. With automated image analysis methods, quantitative and objective measurements are extracted to capture the patient’s condition and disease characteristics in a meaningful and reproducible way. The proposed CDS systems have been validated using data collected from routine practice and clinical trials. The datasets used in this dissertation are from multiple medial centers, which increases the generalizability of the proposed frameworks and trained models. This work aims to research the capacity of AI models toward fully automated CDS systems that can replicate expert judgment and provide insight for the patient. Efforts have been made to improve the generalizability and interpretability of AI/ML models, which are the major limitations that hinder a broad application of AI techniques in practice. The proposed algorithms and strategies in this dissertation leverage big data to improve the healthcare system and disease research. Additionally, the proposed methods are transferable beyond the target application. The contributions of this dissertation have a meaningful impact on applying AI-based systems to clinical and research practice.