Challenging the Binary Classification Paradigm in Longitudinal Health Data Settings

Abstract

Machine learning techniques are commonly used in medical applications for prognostic predictions. However, whereas these are inherently longitudinal problems, a binary classification paradigm currently dominates the literature, in which the disease worsening rate over a predefined time horizon is predicted. If this binary classification approach implies excluding censored instances from the dataset, this introduces a bias into the learning problem, and a survival analysis approach would be more appropriate. Hence, in this paper, we compare the classification performance of the two approaches (binary classification and survival analysis) on the binarized survival analysis target for 69 benchmark survival analysis datasets. Specifically, we focus on decision tree and random forest models, as they have very similar algorithms for both classification and survival analysis. We conclude that, on the binary classification task, the performance of the survival analysis alternative is not statistically significantly different. However, the survival analysis approach allows for a comprehensive prediction of risk profiles over time, while the binary classification approach is limited to a cross-sectional analysis. Thus, we recommend a survival analysis approach when appropriate in longitudinal health data settings, as it increases the expressiveness of the predictions without loss of performance.