A Comparison of Methods for Disease Progression Prediction Through a GoDARTS Study

Abstract

Background: In recent years, a variety of new machine learning methods are being employed in prediction of disease progression, e.g. random forest or neural networks, but how do they compare to and are they direct substitutes for the more traditional statistical methods like the Cox proportional hazards model? In this paper, we compare three of the most commonly used approaches to model prediction of disease progression. We consider a type 2 diabetes case from a cohort-based population in Tayside, UK. In this study, the time until a patient goes onto insulin treatment is of interest; in particular discriminating between slow and fast progression. This means that we are both interested in the results as a raw time-to-insulin prediction but also in a dichotomized outcome making the prediction a classification.Methods: Three different methods for prediction are considered: A Cox proportional hazards model, random forest for survival data and a neural network on the dichotomized outcome. The performance is evaluated using survival performance measures (concordance indices and the integrated Brier score) and using the accuracy, sensitivity, specificity, and Matthews correlation.coefficient for the corresponding classification problems.Results: We found no improvement when using the conditional inference forest over the Cox model. The neural network out performed the conditional inference forest in the classification problem. We discuss the limitations of the three approaches and where they each excel in terms of prediction performance, interpretation, and how they handle data imbalance.