AI Prototype for Streamlining Medical Documentation and Automated Report Generation

  • Julia Schober

Student thesis: Master's Thesis

Abstract

This thesis explores the design and implementation of an artificial intelligence (AI) prototype aimed at automatically identifying hierarchical structures within an eye database,
the Eye Clinic Manager (ECM), developed by INNOFORCE. Specifically developed for
the University Clinic of Ophthalmology and Optometry in Salzburg, the prototype uses
current medical histories from the general outpatient eye clinic to make predictions
about criteria that would otherwise be selected manually. This approach aims to reduce
the documentation burden on medical staff, thereby saving valuable time and costs.
The development of the AI prototype involved conceptualizing the design, breaking
down the overall task into manageable steps, and successfully addressing challenges
in binary text classification. A comprehensive design was created for integrating the
prototype into the ECM, with a thorough analysis of the required steps and technical
framework conditions.
The implementation of the AI prototype began with data conversions and preprocessing
into a machine learning (ML)-friendly format. Several classic ML methods were applied
and evaluated, including logistic regression as a baseline model, support vector classifier (SVC), random forest, and k-nearest neighbor (KNN). A selection of these models
were optimized through hyperparameter tuning to improve performance. Advanced techniques, including artificial neural networks (ANN) and large language models (LLM),
were also designed and utilized to explore their potential for achieving even greater
results.
The findings indicate that while simpler approaches, such as SVC, frequently yielded
satisfactory outcomes, other traditional techniques like decision tree and random forest struggled to accurately capture the underlying patterns in the data, resulting in
sub-optimal predictions of the target criteria. In contrast, diverse ANN architectures
produced moderately successful results. However, these were surpassed by other methods in terms of performance. Notably, LLMs demonstrated the most promising results,
despite limitations in fine-tuning with the full dataset. Overall, several criteria have already been effectively predicted using various techniques, achieving an accuracy in the
range of 85-90%, highlighting the potential for further refinement and optimization.
This thesis assesses the quantitative impact of the AI prototype on reducing documentation workload and examines how different AI approaches and model architectures
influence the accuracy and reliability of predicting hierarchical structures within the
ECM.
Date of Award2024
Original languageEnglish (American)
SupervisorViktoria Dorfer (Supervisor), Herbert Reitsamer (Supervisor) & Christoph Wille (Supervisor)

Cite this

'