Towards automated ideal treatment pathways in healthcare

Details

Supervisors

Absil, Pierre-Antoine ; Winandy, Charles-Eric

Degree label

Master [120] : ingénieur civil en science des données, à finalité spécialisée

Abstract

This master thesis has the objective of helping medical experts in decision making. The concrete application studied is the moveUP index which aims at summarizing in one number on a daily basis the rehabilitation status of a patient who underwent a hip or knee arthroplasty. Firstly, a summary of the existing literature tackling the rehabilitation trajectories of patients undergoing an arthroplasty, their profiling and the prediction of the surgery outcome is presented in this paper. This summary demonstrates that it is possible to draw recovery pathways for hip and knee patients after an arthroplasty. It also shows that the main improvements happen during the six first weeks of the rehabilitation. Searchers disagree on whether it is possible to predict the outcome of a surgery based on pre-surgery data. It was interesting to discover that the methodology used for the pathways creation and the outcome prediction were mainly linear regressions while the profiling of the patients relies on medical scores thresholds. Then, this work proposes a data-based approach to define both patients’ profiles and recovery pathways. The data preprocessing including data cleaning and management of the missing values (through a k-nearest neighbors’ model) are described in detail. A theoretical summary reminds how the machine learning models used in this master thesis work. Thus, the k-means clustering algorithm, the decision tree classifier, the k-nearest neighbors algorithm and the multilayer perceptron (neural network) are explained. Those reminders allow to understand the model implementation with the definition of the hyper parameters and the choice of the best model. Then, the development of patients profiling models is presented. It allows to conclude that the pre-surgery profile of a patient, coherent with the moveUP indexes values, with other pre-surgery data can predict the post-surgery profile of a patient. However, this post-surgery profile does not correspond to a post-surgery moveUP indexes valued range. Those models can thus help for the patient’s selection for surgery. Nevertheless, the post-surgery moveUP indexes formulas need to be improved to be more coherent with the patients post-surgery profiles. Finally, daily moveUP indexes models are developed for each subdimension (symptoms, pain, activity of daily living and quality of live) and for each limb (hip and knee patients separately) for the first six weeks following the surgery with multilayer perceptrons. While the Mean Absolute Error of the models remains high (around 10.75 points), the results obtained are coherent with the literature and the medical experts’ opinion. However, the prediction of the daily moveUP indexes values for the following future days computed using a simple k-nearest neighbors model did not brought good results. In brief, this master thesis demonstrates that (1) the existing pre-surgery moveUP indexes definitions correspond to the patients profiling but the post-surgery moveUP indexes formulas could be improved; (2) the pre-surgery patients characteristics have an impact on the post-surgery results; (3) the post-surgery profile of a patient can be predicted based pre-surgery data with more than 70% of accuracy. (4) Daily moveUP indexes have been developed and show results coherent with the PTs opinion and the literature.