Machine learning for randomised controlled trials: identifying treatment effect heterogeneity with strict control of type I error

BackgroundIt is widely acknowledged that retrospective exploratory analyses of randomised controlled trials (RCTs) seeking to identify treatment effect heterogeneity (TEH) are prone to bias and false positives. Yet the increasing availability of multiple data modalities on subjects and the desire to learn all we can from trial participants motivates the inclusion of such analyses within RCTs. Coupled to this, widespread advances in AI and machine learning (ML) methods hold great potential to utilise such data to characterise subjects exhibiting heterogeneous treatment response.\n\nMethodsWe present new learning strategies for RCT ML discovery methods that ensure strict control of the false positive reporting rate at a pre-specified level. Our approach uses randomised data partitioning and statistical or ML based prediction on held-out data. This can test for both crossover and non-crossover TEH. The former is done via a two-sample hypothesis test measuring overall predictive performance of the ML method. The latter is done via stacking the ML predictors alongside a classical statistical model to formally test the added benefit of the ML algorithm. An adaptation of recent statistical theory allows for the construction of a valid aggregate p-value. This learning strategy is agnostic to the choice of ML method.\n\nResultsWe demonstrate our approach with a re-analysis of the SEAQUAMAT trial. We find no evidence for any crossover subgroup who would benefit from a change in treatment from the current standard-of-care, artesunate, but strong evidence for significant noncrossover TEH within the artesunate treatment group. We find that artesunate provides a differential benefit to patients with high numbers of circulating ring stage parasites.\n\nConclusionsOur ML approach combined with the use of computational notebooks and version control can improve the robustness and transparency of RCT exploratory analyses. The methods allow researchers to apply the latest ML techniques safe in the knowledge that any declared associations are statistically significant at a user defined level.