A framework for confounder considerations in AI-driven precision medicine

Abstract

AbstractIntroductionArtificial intelligence holds promise for individualized medicine. Yet, transitioning models from prototyping to clinical applications poses challenges, with confounders being a significant hurdle. We introduce a two-dimensional confounder framework (Confound Continuum), integrating a statistical dimension with a biomedical perspective. Informed and context-sensitive confounder decisions are indispensable for accurate model building, rigorous evaluation and valid interpretation.MethodsUsing prediction of hand grip strength (HGS) from neuroimaging-derived features in a large sample as an example task, we develop a conceptual framework for confounder considerations and integrate it with an exemplary statistical investigation of 130 candidate confounders. We underline the necessity for conceptual considerations by predicting HGS with varying confound removal scenarios, neuroimaging derived features and machine learning algorithms. We use the confounders alone as features or together with grey matter volume to dissect the contribution of the two signal sources.ResultsThe conceptual confounder framework distinguishes betweenhigh-performancemodels andpure linkmodels that aim to deepen our understanding of feature-target relationships. The biological attributes of different confounders can overlap to varying degrees with those of the predictive problem space, making the development ofpure linkmodels increasingly challenging with greater overlap. The degree of biological overlap allows to sort potential confounders on a conceptualConfound Continuum. This conceptual continuum complements statistical investigations with biomedical domain-knowledge, represented as an orthogonal two-dimensional grid.Exemplary HGS predictions highlighted the substantial impact of confounders on predictive performance. In contrast, choice of features or learning algorithms had considerably smaller influences. Notably, models using confounders as features often outperformed models relying solely on neuroimaging features.ConclusionOur study provides a confounder framework that combines a statistical perspective on confounders and a biomedical perspective. It stresses the importance of domain expertise in predictive modelling for critical and deliberate interpretation and employment of predictive models in biomedical applications and research.Short descriptionThe paper explores the challenges of transitioning predictive models from scientific prototyping to clinical use, with a focus on the significant impact of confounders. Using the example of predicting hand grip strength in the UK Biobank, the study introduces a framework that integrates statistical and biomedical perspectives on confounders, emphasizing the vital role of informed confounder decisions for accurate model development, evaluation and interpretation.