Unlocking the clinical potential of AI

Artificial intelligence (AI) is a powerful and rapidly developing technology that is fast becoming part of everyday society. It has enormous potential to enhance personalised medicine and improve human health. However, despite its growth, it has not yet become a mainstay in clinical practice.

In a comment piece published in Nature Medicine , Ben MacArthur , professor of quantitative biomedicine in the Faculty of Medicine and the School of Mathematical Sciences alongside colleagues from the Alan Turing Institute, discusses why that is and what is needed to unlock the clinical potential of AI.

AI models conventionally focus on maximising predictive accuracy, but this approach falls often short in the clinical context, since it does not properly account for the fact that each patient is a unique individual or properly quantify and communicate the uncertainty associated with clinical AI decisions.

The paper says: “The ability of clinicians to convey uncertainties to patients with sensitivity is central to clinical medicine. Taking AI tools into the clinic will therefore require new approaches to model development that use clinically relevant metrics, recognize the distinctiveness of each individual patient and provide personalized measures of performance. Successful clinical AI tools cannot simply maximize predictive accuracy; they must also convey uncertainty.”

The research paper considers the concept of "personalised measures of uncertainty", which involves distinguishing between two types of uncertainty: epistemic (knowledge-based) uncertainty, which can be reduced with more clinical knowledge, and aleatoric (data-based) uncertainty, arising from natural random variations in measurements. By better quantifying these uncertainties, AI can provide more accurate assessments.

The paper also highlights how a new set of tools called conformal prediction can provide a better understanding of uncertainty for each patient. Conformal prediction allows AI models provide to a list of personalised possible diagnoses for the patient, which the clinician can follow up on using their expertise to make a tailored decision for the care.

The authors conclude by saying: “Widespread use of AI in the clinic will require AI tools to move away from simply maximizing predictive accuracy toward harnessing uncertainty and supporting clinicians in making well-informed decisions for each individual patient. New methods such as conformal prediction are making this transition possible and will be critical to truly integrating AI into clinical practice, helping doctors improve healthcare for all patients.”