The authors used a clinical prediction model to estimate the individual risk of developing breast cancer, but it is questionable if models can do this. A recently updated description of these models reads (1): An individuals’ risk of an event as estimated by a prediction model refers to the probability of that event in the subgroup of individuals with the same predictor values. Truly individual risks do not exist because of the ‘reference class problem’: an individual can belong to an infinite number of subgroups. By choosing a different set of predictors to condition on, the reference class changes, and so does the risk.

It may seem paradoxical that using prediction models for decision making can benefit    outcomes at a population level without having trustworthy risk estimates for individuals    from that population. ... Clinical utility at the population level, for example in terms of     cost-effectiveness, is sufficient to warrant the use of a decision strategy based on a  model. We may need to accept that two models may have the same clinical utility yet     very different risk estimates for the same individual, often to the extent that model A     recommends treatment initiation, but model B does not.

Another concluded that (2):

clinical prediction models cannot, do not, and need not estimate individual risk.

Since there is no evidence individual risk exist (3) and clinical prediction models don't estimate individual risk, these models are better understood to be risk stratifying a population. They separate a population into groups differing in risk, not individuals differing in risk, which can then be used to devise policies that efficiently allocate surveillance or preventive measures. This may make sense for the general population, but is unlikely to be useful for BRCA1 and BRCA2 pathogenic variant carriers. If predictors are added to a model, the risk distribution often broadens (detected by measures of discrimination) which may allow even more efficient allocations. This was nicely shown by the figure depicting the broader risk distribution resulting from addition of the polygenic risk score to age alone.

Discordance in predictions for an individual across models is well documented and has been called "the multiverse of madness (4), having been observed with breast cancer risk models (5) and polygenic risk models (6). This does not impact their use for population risk stratification, but should preclude their use for individual risk estimation.

Holmberg and Parascandola critiqued the use of breast cancer risk models, including the presentation of numbers as personalized risks (7): But we would argue that the more general problem here is with the promise of individualised risk information, which implies a sense of authority and certainty about the individual. Framing risk estimates as individualised may be misleading as it implies a high level of specificity for the individual. The use of individualised risk estimates suggests that a risk model says something personal about a specific individual rather than an entire group.

Models including more or better predictors or using machine learning may provide greater discrimination but will not evade the reference class problem. So, this problem will persist and geneticists using clinical prediction models need to understand this counterargument to the conventional narrative on individual risk. To date, the reference class problem has been a well-kept secret in medicine with only 10 results in PubMed. A new points to consider statement from the ACMG (8) does caution that risk estimates cannot be directly applied to an individual and that discordant risks may be estimated. Nevertheless, personalized or individualized risk estimates remain the goal (with future model improvements expected to make this possible) and the discordance is not flagged, as it usually is, as a reason to question the use of these numbers. There no discussion of the reference class problem and its implication that individual risks don't exist and that the continuum of risk applies to groups, not individuals, so models provide population risk stratification suitable for policy development but not for counseling.

As geneticists considering counseling based on a clinical prediction model need to know about the reference class problem and its implications, this merits inclusion in the discussion of limitations when the research is eventually published. Ralph Stern Division of Cardiovascular Medicine Department of Internal Medicine University of Michigan

1. Barreñada L, Steyerberg EW, Timmerman D, Thomassen D, Wynants L, Van Calster B. The fundamental problem of risk prediction for individuals: health AI, uncertainty, and personalized medicine. arXiv. Preprint posted online 2025. doi:10.48550/ARXIV.2506.17141

2. Stern RH. Accuracy of Preoperative Risk Assessment: Individuals versus Groups. Am J Med. 2025;138(6):926-927. doi:10.1016/j.amjmed.2025.02.007

3. Dawid P. On individual risk. Synthese. 2017;194(9):3445-3474. doi:10.1007/s11229-015-0953-4

4. Riley RD, Pate A, Dhiman P, Archer L, Martin GP, Collins GS. Clinical prediction models and the multiverse of madness. BMC Med. 2023;21(1):502. doi:10.1186/s12916-023-03212-y

5. Paige JS, Lee CI, Wang PC, et al. Variability Among Breast Cancer Risk Classification Models When Applied at the Level of the Individual Woman. J Gen Intern Med. 2023;38(11):2584-2592. doi:10.1007/s11606-023-08043-4

6. Clifton L, Collister JA, Liu X, Littlejohns TJ, Hunter DJ. Assessing agreement between different polygenic risk scores in the UK Biobank. Sci Rep. 2022;12(1):12812. doi:10.1038/s41598-022-17012-6

7. Holmberg C, Parascandola M. Individualised risk estimation and the nature of prevention. Health, Risk & Society. 2010;12(5):441-452. doi:10.1080/13698575.2010.508835 1.

8. Pal T, Christopher J, Astiazaran-Symonds E, et al. Consideration of inherited cancer risk on a continuum: An international and multidisciplinary perspective: A points to consider statement of the American College of Medical Genetics and Genomics (ACMG). Genet Med. 2026;28(3):101659. doi:10.1016/j.gim.2025.101659