AI-Driven Approaches for Biomedical Data Analysis and Clinical Decision Support Systems: From Diagnosis to Prevention

Machine learning and artificial intelligence have become transformative tools in healthcare, providing novel solutions to complex biomedical challenges. The increasing availability of large-scale, multi-modal medical data enables AI-driven approaches to improve diagnostics, treatment planning, and patient monitoring. However, real-world medical data often suffer from high dimensionality, heterogeneity, and limited annotation, posing significant challenges to AI model development. This dissertation addresses these challenges by developing and applying machine learning methodologies to four key areas of biomedical research: 1) inhalation injury grading with bronchoscopy imaging, 2) predictive modeling of fall emergency occurrence in smart home environments, 3) smartphone-based keratoconus detection, and 4) automated measurement of total burn surface area (TBSA). In the first study, I propose a novel deep learning-driven grading system for inhalation injuries, incorporating bronchoscopy images and patient ventilation duration as an objective severity indicator. The system utilizes generative data augmentation techniques and transfer learning to address dataset limitations, significantly improving classification accuracy and clinical applicability. The second study introduces an interpretable, multi-modal AI model for predicting fall emergencies in older adults. By integrating static health information with time-series movement data, this model identifies high-risk individuals and provides actionable insights into fall prevention strategies. The third study presents a smartphone-based keratoconus detection method that captures side-view Placido disc reflections to diagnose and localize corneal abnormalities without requiring specialized clinical equipment. This approach enables cost-effective and accessible keratoconus screening. Finally, I developed an AI-driven segmentation framework for TBSA measurement, addressing the limitations of existing burn assessment methods. By leveraging deep learning-based semantic segmentation techniques, the proposed model accurately identifies burned regions, providing clinicians with an automated, objective assessment of burn severity. The contributions of this dissertation include novel AI-driven methodologies tailored to specific biomedical challenges, integration of multi-modal data to improve prediction accuracy, and development of accessible and interpretable machine learning models for clinical applications. These studies collectively advance the role of AI in medical imaging, patient monitoring, and treatment decision-making, contributing to improved healthcare outcomes and broader adoption of machine learning in clinical practice.