Predicting Wound Healing Using Ordinal Classification on Medical Record Data
Keywords
Loading...
Authors
Issue Date
2022-02-04
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
In the Netherlands, many people suffer from complex wounds, that take a long time to heal and require specialized care. Early prediction of such complex wounds is a difficult task, even for wound care experts, because the wound healing process is highly complex and affected by many factors, including wound and patient characteristics. In this thesis, factors that can affect wound healing were identified based on literature review and by consulting experts. Based on this list of features, data was collected which was then used to perform wound healing duration prediction for different types of wounds. Ordinal classification methods were compared to binary classification baselines, to evaluate whether ordinal classification methods can more accurately predict wound healing duration for a larger number of classes. Furthermore, performance was not only compared to a simple baseline, but to expert performance as well. Binary classification using XGBoost reached an accuracy of 68%, compared to an accuracy of 74% for wound experts. When doing multiclass classification, there is no difference between model and expert predictions. The addition of feature vectors extracted from a neural network trained on wound tissue segmentation seemed to be relevant for accurately predicting wound healing duration, as it was one of the most important features according to model feature importance. As automatic wound healing prediction approaches expert level performances, opportunities arise to use such algorithms to perform early filtering of patients at risk of complex wounds without having to consult experts, allowing for earlier interventions, which could hopefully speed up the wound healing process.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen