Comparison of machine learning clustering algorithms for detecting heterogeneity of treatment effect in acute respiratory distress syndrome: A secondary analysis of three randomised controlled trials

BackgroundHeterogeneity in Acute Respiratory Distress Syndrome (ARDS), as a consequence of its non-specific definition, has led to a multitude of negative randomised controlled trials (RCTs). Investigators have sought to identify heterogeneity of treatment effect (HTE) in RCTs using clustering algorithms. We evaluated the proficiency of several commonly-used machine-learning algorithms to identify clusters where HTE may be detected.MethodsFive unsupervised: Latent class analysis (LCA), K-means, partition around medoids, hierarchical, and spectral clustering; and four supervised algorithms: model-based recursive partitioning, Causal Forest (CF), and X-learner with Random Forest (XL-RF) and Bayesian Additive Regression Trees were individually applied to three prior ARDS RCTs. Clinical data and research protein biomarkers were used as partitioning variables, with the latter excluded for secondary analyses. For a clustering schema, HTE was evaluated based on the interaction term of treatment group and cluster with day-90 mortality as the dependent variable.FindingsNo single algorithm identified clusters with significant HTE in all three trials. LCA, XL-RF, and CF identified HTE most frequently (2/3 RCTs). Important partitioning variables in the unsupervised approaches were consistent across algorithms and RCTs. In supervised models, important partitioning variables varied between algorithms and across RCTs. In algorithms where clusters demonstrated HTE in the same trial, patients frequently interchanged clusters from treatment-benefit to treatment-harm clusters across algorithms. LCA aside, results from all other algorithms were subject to significant alteration in cluster composition and HTE with random seed change. Removing research biomarkers as partitioning variables greatly reduced the chances of detecting HTE across all algorithms.InterpretationMachine-learning algorithms were inconsistent in their abilities to identify clusters with significant HTE. Protein biomarkers were essential in identifying clusters with HTE. Investigations using machine-learning approaches to identify clusters to seek HTE require cautious interpretation.FundingNIGMS R35 GM142992 (PS), NHLBI R35 HL140026 (CSC); NIGMS R01 GM123193, Department of Defense W81XWH-21-1-0009, NIA R21 AG068720, NIDA R01 DA051464 (MMC)