A framework for confounder considerations in AI-driven precision medicine

Machine learning (ML) models hold promise in precision medicine by enabling personalized predictions based on high-dimensional biomedical data. Yet, transitioning models from prototyping to clinical applications poses challenges, with confounders being a significant hurdle by undermining the reliability, generalizability, and interpretability of ML models. Using hand grip strength (HGS) prediction from neuroimaging data from the UK Biobank as a case study, we demonstrate that confounder adjustment can have a greater impact on model performance than changes in features or algorithms. An ubiquitous and necessary approach to confounding is by statistical means. However, a pure statistical viewpoint overlooks the biomedical relevance of candidate confounders, i.e. their biological link and conceptual similarity to actual variables of interest. Problematically, this can lead to biomedically not-meaningful confounder-adjustment, which limits the usefulness of resulting models, both in terms of biological insights and clinical applicability. To address this, we propose a two-dimensional framework, the Confound Continuum, that combines both statistical association and biomedical relevance, i.e. conceptual similarity, of a candidate confounder. The evaluation of conceptual similarity assesses on a continuum how much two variables overlap in their biological meaning, ranging from negligible links to expressing the same underlying biology. It thereby acknowledges the gradual nature of the biological link between candidate confounders and a predictive task. Our framework aims to create awareness for the imperative need to complement statistical confounder considerations with biomedical, conceptual domain knowledge (without going into causal considerations) and thereby offers a means to arrive at meaningful and informed confounder decisions. The position of a candidate confoudner in the two-dimensional grid of the Confound Continuum can support informed and context-specific confounder decisions and thereby not only enhance biomedical validity of predictions but also support translation of predictive models into clinical practice.