Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches

Dear Editor

Congratulations to Riley et colleagues for highlighting the importance of presenting uncertainty around risk estimates provided by clinical prediction models (1). As nicely reviewed by the authors, there are numerous causes of uncertainty and bias due to poor methodological standards and inappropriate risk prediction model development techniques, poor model stability, or insufficient external evaluation.

A deeper reason of overlooking uncertainty in prediction models stems from the confusion between risk and disease. As argued by Aronowitz, "one underappreciated consequence of modern clinical and public health practices is that the experience of being at risk for disease has been converging with the experience of disease itself" (2). This conflation between risk and disease is the result of increased biological, clinical, and epidemiological knowledge on the risk particularly of chronic diseases, and more specifically of the application of a population perspective to individual-level clinical care which feeds the false notion that a prediction accurate at a population or group level could be directly applied to an individual patient.

One harmful effect of ignoring uncertainty in clinical prediction models arises when these predictions are translated into clinical decision rules (3). While risk predictions are intended to guide preventive services and treatments at a group level, it is tempting but deceptive to apply them bluntly to individual patient decisions—on a case-by-case basis—to simplify the inherently complex nature of clinical judgment (3). This reflects the unfulfilled promise of a data-driven personalized medicine trying to overcome the failure of epidemiology in knowing who is and who is not eventually getting the disease (4, 5).

We agree with Riley et al. that improving the communication of uncertainty in risk prediction is essential. However, this should be accompanied by the acknowledgement of the complexity of clinical decision-making, emphasizing that, regardless of the level of risk prediction certainty, randomness always plays a role at the individual level (5).

References
1) Riley RD, Collins GS, Kirton L, Snell KI, Ensor J, Whittle R, Dhiman P, van Smeden M, Liu X, Alderman J, Nirantharakumar K, Manson-Whitton J, Westwood AJ, Cazier JB, Moons KGM, Martin GP, Sperrin M, Denniston AK, Harrell FE Jr, Archer L. Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches. BMJ. 2025; 388:e080749.
2) Aronowitz RA. The converged experience of risk and disease. Milbank Q. 2009 Jun;87(2):417-42.
3) Gérvas J, Starfield B, Heath I. Is clinical prevention better than cure? Lancet. 2008; 372(9654):1997-9.
4) Chiolero A. There is nothing personal. Arch Intern Med. 2012; 172(21):1691-2.
5) Smith GD. Epidemiology, epigenetics and the 'Gloomy Prospect': embracing randomness in population health research and practice. Int J Epidemiol. 2011; 40(3):537-62.