Categorizing Medical Equipment Issues and Solutions Statements Using AI/ML

Abstract

Categorizing unstructured free text is a laborious task and requires someone with expertise in the respective field. Our study leverages SoftPro Medical Solutions’ extensive database of free-text documents. The database contains authentic problem descriptions about medical equipment, written by technicians and customers from SoftPro’s clients. The thesis aims to explore multiple unsupervised and supervised machine learning algorithms and lay the foundations to categorize future problem descriptions, in order for SoftPro to integrate this into their system at a later stage. Various machine learning algorithms and pre-processing techniques are evaluated to learn the weaknesses and strengths of each algorithm in order to select viable options for the problem at hand. The unsupervised learning algorithms are measured using the metrics Silhouette Score, Davies Bouldin Index, and Calinski Harabasz Index. While the supervised learning algorithms are measured with an F1 Score and the percentage in the form of a fraction of total correct classifications. Our findings suggest that among our researched algorithms the ones most suitable were the Support Vector Machine and Naive Bayes, with Naive Bayes achieving the better results out of the supervised learning algorithms. For unsupervised learning, it was decided to approach density-based clustering algorithms DBScan and OPTICS because of their resistance to outliers. However, further testing would be required to determine the better algorithm of the two.

Date of Award	2023-Jun
Original language	English
Supervisor	Daniel Einarson (Supervisor) & Ola Johansson (Examiner)