All Times EDT
Keywords: machine learning, clinical trial, explainable AI
In an era of ever-increasing amounts of patient data measured in interventional clinical trials, combined with a need for precision medicine therapies, clinical statistics as well is shifting towards greater appreciation and accommodation for the true biological complexity of human beings and patients as such. This system complexity is characterized by higher-order interactions and potentially non-linear and non-monotonous associations between essentially every system component such as treatment, baseline characteristics, demography, lab-data, medical history, concomitant medications and many more. To accommodate for such complexity, to quantify it and to enable insights regarding the manifold characteristics that may determine patient-specific treatment benefit, machine learning (ML) based methodology is most suited. Yet, modelling complexity always comes with a trade-off pertaining to results interpretability. Ameliorating this, recent years have seen a tremendous development of methods coined “explainable artificial intelligence” (XAI) and are now at a level that make ML results amenable to life scientists and physicians as key decision makers in drug development. At Bayer AG Clinical Development we have developed an industrialized software solution to efficiently compute complex ML models trained with literally all data collected in specific late phase trials. By using XAI were able to detect, explain and interpret said highly complex patient prognostic and treatment-modifying characteristics retrospectively in a growing number of phIII trials and were successful in communicating this type of supportive evidence internally. We propose that such ML based analyses results of late phase interventional trials may become valuable supportive evidence in (more personalized) drug development both for decision making within a company as well as for discussion with regulators and the academic community.