Good things come in threes: evaluating clinical utility of machine learning-derived clusters

Background Machine learning (ML)-based cluster analysis is a common method for subtyping medical conditions and presentations of disease states. Recent advancements in algorithms and increasing access to vast healthcare data have further increased the use of such ML models. However, practical implementation is lacking, largely due to methodological limitations and insufficient reporting and clinical contextualization in the extant literature, with no existing guidelines for this purpose.Objective To propose a general framework to assess and ensure clinical utility of ML-derived clusters.Discussion The proposed framework encompasses three domains, focusing on clinical relevance, stability and generalizability, and ease of identification, enabling researchers and readers to ensure clinical utility of ML-generated subgroups.Many diseases and clinical conditions are complex and vary in how they present, develop, and respond to treatment. Understanding these variations across individuals, also called phenotypes or subtypes of disorders, is important to offer personalized treatment alternatives and improve long-term outcomes. Machine learning is increasingly used in these efforts, as these techniques can greatly facilitate exploration of complex clinical data. One of the most common methods for this is cluster analysis, in which an algorithm, more or less independently, can suggest subgroups (clusters) from unlabeled data. Despite promising findings, cluster analyses remain abstract and difficult to implement into clinical practice. Much of this is due to machine learning algorithms largely being so-called "black boxes," for which it is difficult to understand how the output was obtained. Furthermore, studies using cluster analyses rarely report how they performed their work in sufficient detail and often miss important considerations for evaluating clinical utility of their results. In this work, we propose an easy-to-use framework, to be used by readers and researchers alike, for ensuring clinical utility of machine learning-derived clinical subgroups.