Predictive Modelling of Pain Based on Quantitative Markers Extracted From MRI Scans of Knees

Specialeforsvar ved Michelle Bruhn Jensen

Titel:  Predictive Modelling of Pain Based on Quantitative Markers Extracted From MRI Scans of Knees

Abstract: The purpose of this study was to investigate if a predictive model could be used to predict the pain when people are known to be suffering from knee osteoarthritis (OA), as well as which predictive model that would predict most accurately, so when a new patient is diagnosed with knee OA from their symptoms, then a more accurate pain score can be achieved, so known actions in reducing the pain could be made in regard of the patient. Since the dependent variable is numerical this thesis uses different regression models in order to investigate the model best suited for the data gathered from the OA Initiative, as well as data provided from Erik Dam with quantitative markers extracted from MRI scans of knees. Data was considered with different uses of cross validation and different predictive models. A predictive model that were convincingly better than the others were identified by the estimated loss of different loss functions. None of the predictive models managed to explain the variation of the WOMAC pain score in a satisfactorily manner, but it was still possible to conclude that random forest regression was better than different linear regression models and support vector regression. As well as K-fold cross validation providing better results than the holdout method.

Vejleder: Anders Tolver
Censor:   Anders Rønn-Nielsen, CBS