Someone who has tested positive for the expanded huntingtin gene lives with an unusual kind of arithmetic. The diagnosis has not been made, may not be made for a decade or more, and yet the question of when hangs over trial recruitment, work decisions and family planning. Researchers answer that question with prediction models: Langbehn's 2004 survival model, the CAG-Age Product (CAP) model, the Multivariate Risk Score (MRS), and the Prognostic Index Normed (PIN) model. Each takes what is known about a person, at minimum the CAG repeat length and the age at enrollment, and turns it into an estimate of the time remaining before a clinician will call motor signs unequivocal. The models do not always agree. Until recently, remarkably little work had tested them side by side on data none of them had seen before.
That is what a preprint from Kyle F. Grosser, Abigail G. Foes, Stellen Li, Vraj Parikh and Tanya P. Garcia of the Department of Biostatistics at the University of North Carolina at Chapel Hill, with Sarah C. Lotspeich of the Department of Statistical Sciences at Wake Forest University, sets out to do. Posted to arXiv on 5 November 2025 and substantially revised on 14 March 2026, it is a methods paper, not a treatment study, and it has not yet been peer reviewed. The authors argue that earlier comparisons of these four models carried two flaws. The first they call double-dipping: judging a model on the same dataset used to build it, which flatters it. The second is statistical and, for families, the more interesting one: the conventional accuracy scores used to rank these models quietly assume that everyone's outcome is known.
In Huntington's research, it is not. The authors note that modern prodromal cohorts run right-censoring rates above 80 percent, meaning more than four in five gene carriers reach the end of follow-up without a clinical diagnosis. Their times to diagnosis are not missing at random and they are not zero; they are simply later than anyone watched for. The contrast with the oldest model is stark. The Langbehn model, the preprint records, was developed from retrospective data at 40 clinical sites across Africa, Asia, Europe and North America, restricted to patients with 41 to 56 CAG repeats: 2,913 people, of whom 2,298 had been diagnosed by the end of the study, leaving a censoring rate of 21 percent. The MRS model, by contrast, came from PREDICT-HD participants with 36 or more CAG repeats who were not yet diagnosed at baseline, an analytic sample of 1,078 in which 225 were diagnosed during the study and 853 were not, a censoring rate of 79 percent. A metric that ignores those 853 people is not measuring what it claims to measure.
So the team applied all four models to Enroll-HD, a dataset used to develop none of them, pulling age at enrollment, CAG repeats, the UHDRS Diagnostic Confidence Level, the UHDRS Total Motor Score, the Symbol Digit Modalities Test and Stroop scores, and restricting the sample to gene-expanded participants who were not yet clinically diagnosed at their first visit. They scored the models with Uno's concordance statistic and a Kaplan-Meier-based ROC analysis, both of which are constructed to handle heavy censoring. The revised version reports that on either metric, and whether the models were run with their originally published parameters or with parameters re-estimated in Enroll-HD, the MRS model, which uses the most variables, performed best. The much simpler PIN model came close while asking less of the clinic, a point the authors press, since extra covariates mean extra tests for the person sitting in the chair.
What follows from that is a trial design argument rather than a clinical one. These models are used for sample enrichment: choosing participants who are close enough to diagnosis that a preventive therapy could plausibly show it has delayed anything within the life of a trial. Choose badly and the trial is too small to answer its own question. The first version of the preprint put it bluntly, saying previous sample size estimates using such models "would lead to underpowered trials". The revised version works through an optimal threshold for deciding who is more likely to be diagnosed, and recalculates how many participants per arm a preventive study would actually need. For a community that has watched more than one Huntington's trial end without a clean answer, the size of the trial is not a footnote.
For a family, none of this changes a prediction that has already been given, and none of it is a new test. It is a comparison of yardsticks, made with data from people who volunteered for a long observational study and then, most of them, went home undiagnosed. The authors are careful to say they measured risk stratification, the ability to put people in the right order from nearest to furthest from diagnosis, and not calibration, the question of whether a model's predicted probability of diagnosis by a given year matches what actually happens. They suggest calibration plots and a censoring-adjusted Brier score as the next step. Ordering people correctly and telling any one of them the right year are different achievements, and only the first was tested here.