JMIR Publications

ABSTRACT

The rapid uptake of supervised machine learning (ML) in clinical prediction modelling, particularly for binary outcomes based on tabular data, has prompted debate about its comparative advantage over traditional statistical logistic regression (LR). While ML has demonstrated superiority in unstructured data domains, its performance gains in structured, tabular clinical datasets remain inconsistent and context-dependent. This viewpoint synthesises recent comparative studies and simulation findings to argue that there is no universal best algorithm, its performance depends heavily on dataset characteristics, such as linearity, sample size, number of candidate predictors, and minority class proportion. We advocate for LR as a baseline model in clinical prediction research due to its interpretability and stability, with ML approaches considered only when they show meaningful improvements across multiple performance metrics, particularly clinical utility. Ultimately, we argue that investments in data quality, not model complexity, are more likely to enhance the reliability and utility of clinical prediction models.