Machine Learning Is Not Just for Prediction: Its Role as an Exploratory Analytical Tool in Medicine

Machine learning (ML) has been primarily used for predictive modeling in medical research, but this reflects only part of its potential. This review proposes a conceptual distinction between predictive ML and exploratory ML. Predictive ML aims to maximize accuracy on unseen data. Exploratory ML focuses on identifying underlying structures in data to generate hypotheses. Exploratory ML plays an important role under conditions where conventional hypothesis-driven statistics have limitations, including high-dimensional data, small sample sizes, and biopsy-inaccessible organs such as the vascular system. Because ML-derived results are based on associations rather than causality, they should be interpreted as hypothesis-generating rather than confirmatory. Methods including unsupervised learning, interpretable supervised learning, and network analysis are discussed as exploratory ML approaches. The differences in objectives and evaluation criteria between exploratory ML and conventional hypothesis-driven statistics are also discussed, together with the structural gap in peer review. The key argument is that studies using exploratory ML should be evaluated not by predictive performance but by the stability, reproducibility, and interpretability of the identified structures. Without this shift, exploratory analyses may be systematically misjudged within current evaluation standards. This perspective may bridge exploratory analysis and confirmatory research and support new study designs in medicine.