Enhancing Patient Understanding of Medical Findings Through NLP and 3D Models

Abstract

This thesis investigates patient communication in clinical settings, recognizing that effective communication is vital for ensuring patient compliance and successful treatment outcomes. Complex clinical jargon and inadequately explained pathological processes often lead to patient confusion and anxiety. Advances have been made in medical text simplification, yet personalized visualizations remain underexplored. Specifically, anatomical models can provide patients with foundational knowledge about the human body, enabling clearer explanations of their medical conditions. The challenge lies in the implicit nature of anatomical references in clinical texts, where discussion focuses mainly on pathology. Anatomical entities are connected to pathological processes, but often not mentioned explicitly, for example a report on a heart attack may mention "myocardial infarction", without mentioning the associated "heart muscle". This thesis compares existing NER methods in their capability to extract implicit entities, and introduces a novel pipeline that leverages foundational biomedical entity relationships from the Unified Medical Language System (UMLS), a compendium of controlled vocabularies that provides structured mappings of relationships between biomedical entities. Extracted entities are visualized through a user interface for a 3D anatomical model, guided by a custom multi-parameter algorithm that optimizes context and clarity. Parameters include contextual distance to surrounding structures, as well as techniques for dimming, highlighting and adjusting opacity. These visualization parameters proved effective in enhancing visual representations by emphasizing relevant structures and minimizing visual clutter. An analysis on the inclusion radius of surrounding structures revealed diminishing returns for all organs tested. A custom camera positioning algorithm was used to automatically center and orient the viewpoint based on the anatomical target’s bounds; this approach effectively improved the clarity and framing of visualizations. To assess annotation quality, large language models were employed as automated evaluators, scoring outputs on a five-point scale. A dedicated validation experiment demonstrated that these models could reliably distinguish between expert-curated and nonsensical annotations, supporting their use as scalable, reproducible evaluation tools. Results show that the proposed method outperforms baselines in annotation quality, with statistically significant and practically meaningful improvements. While opportunities for refinement remain, this research lays the foundation for broader applications in scenarios requiring the extraction and visualization of implicit biomedical entities.