Abstracts Category : Other

Predicting waiting times in radiation oncology using machine learning

by Ackeem Joseph

This thesis describes a study of machine learning and its application to waiting times in radiation oncology. Specifically, an evaluation of waiting time estimates for daily radiation treatment appointments at the McGill University Health Centre was conducted and two unique communication tools for conveying these waiting time estimates were developed. To evaluate waiting time estimates, a subset of previously-treated patient records was used for training and modelling by several off-the-shelf machine learning algorithms and tested using treatment records unseen by the algorithms. Furthermore, machine learning was explicitly used to predict treatment durations in order to calculate an overall waiting time estimate. It was found that the random forest regression algorithm provided the best model for daily radiation treatment durations. On average, the model is able to predict treatment durations to within 7.2 minutes. This is a significant improvement on the rough estimates currently given to patients, and also on the estimates one would obtain by conducting simple averages on all patients. One of the tools developed to communicate waiting time estimates to patients made use of the best-trained model by simulating the delivery of waiting time estimates to all patients in real-time. Waiting time estimates from this study have not yet been delivered to patients. However, a separate and ongoing research is discussed in this dissertation, examining the delivery of personal health information, including waiting time estimates, to patients via a mobile phone application. Ce mmoire dcrit une tude de l'apprentissage machine et son application aux temps d'attente en radio-oncologie. Plus prcisment, le temps d'attente des rendez-vous quotidiens des traitements externes au Centre universitaire de sant McGill ont t estim et deux outils de communication uniques pour transmettre ces estimations ont t labores. Un sous-ensemble de patients traits prcdemment a t utilis pour la formation et la modlisation l'aide de plusieurs algorithmes d'apprentissage machine. Ceux-ci ont t tests en utilisant les donnes exclues par la modlisation. De plus, l'apprentissage machine a t utilis explicitement pour prvoir les dures de traitement afin de calculer une estimation globale du temps d'attente. L'algorithme de random forest regression fournissait le meilleur modle pour les dures quotidiennes de traitement par irradiation. En moyenne, le modle est capable de prdire la dure des traitements a l'intrieur de 7,2 minutes. Il s'agit d'une amlioration considrable par rapport l'stimation du temps d'attente actuellement transmis aux patients et aussi par rapport aux estimations que l'on obtiendrait en effectuant des moyennes simples. Un des outils pour communiquer le temps d'attente simule la livraison des donnes tous les patients en temps rel en utilisant le meilleur modle d'apprentissage machine. ce jour, les estimations de temps d'attente de cette tude n'ont pas encore t fournies aux patients.Advisors/Committee Members: John Kildea (Internal/Supervisor).