The medical AI insurgency: what physicians must know about data to practice with intelligent machines

Abstract Machine learning (ML) and its parent technology trend, artificial intelligence (AI), are deriving novel insights from ever larger and more complex datasets. Efficient and accurate AI analytics require fastidious data science—the careful curating of knowledge representations in databases, decomposition of data matrices to reduce dimensionality, and preprocessing of datasets to mitigate the confounding effects of messy (i.e., missing, redundant, and outlier) data. Messier, bigger and more dynamic medical datasets create the potential for ML computing systems querying databases to draw erroneous data inferences, portending real-world human health consequences. High-dimensional medical datasets can be static or dynamic. For example, principal component analysis (PCA) used within R computing packages can speed & scale disease association analytics for deriving polygenic risk scores from static gene-expression microarrays. Robust PCA of k -dimensional subspace data accelerates image acquisition and reconstruction of dynamic 4-D magnetic resonance imaging studies, enhancing tracking of organ physiology, tissue relaxation parameters, and contrast agent effects. Unlike other data-dense business and scientific sectors, medical AI users must be aware that input data quality limitations can have health implications, potentially reducing analytic model accuracy for predicting clinical disease risks and patient outcomes. As AI technologies find more health applications, physicians should contribute their health domain expertize to rules-/ML-based computer system development, inform input data provenance and recognize the importance of data preprocessing quality assurance before interpreting the clinical implications of intelligent machine outputs to patients.