One-Year Target Lesion Revascularization Prediction Shows Uncertain Incremental Accuracy After Competing-Death Exclusion [Letter] Authors: Jiaqing Liu Published: 2026 Dear editor Derks et al developed models for one-year target lesion revascularization (TLR) after percutaneous coronary intervention (PCI) with drug-eluting stents. The nationwide main cohort included 96,758 patients, and a procedurally enriched subcohort included 11,789 patients. The reported apparent mean Brier scores were 0.021 and 0.024, respectively.1 However, calculations from the published event counts show that both values closely approximate the error from assigning every analysed patient the cohort event rate. The authors acknowledged modest discrimination but described the models as having strong accuracy and calibration. The aggregate calculation does not show that the models contain no predictive information. It shows that the published precision cannot establish material incremental accuracy. A separate eligibility decision, which removed patients who died within one year without TLR, further limits interpretation at the index PCI. TLR occurred in 2099 patients in the main cohort and 289 in the subcohort. A constant prediction equal to each cohort’s event proportion has a Brier score of p(1-p). This produces null benchmarks of 0.021223 and 0.023913. Using the printed model values as midpoints gives scaled Brier scores of 1.05% and −0.36%, calculated as one minus the model-to-null Brier ratio.2 The printed scores have only three decimal places, so compatible unrounded values are more informative than these midpoint calculations. They yield scaled Brier ranges of −1.31% to 3.41% in the main cohort and −2.45% to 1.73% in the subcohort. These ranges include no improvement and performance below the constant-risk benchmark. Consequently, the main article does not demonstrate a material reduction in overall prediction error. Unrounded raw, null, and scaled Brier scores with uncertainty would resolve this question. Calibration evidence reported in the main article also rests on Hosmer-Lemeshow P-value ranges and their arithmetic means across imputations. Such P values neither quantify the direction nor the magnitude of miscalibration, and non-rejection does not establish close agreement between predicted and observed risks.3 Calibration intercepts, slopes, smooth curves, and uncertainty would provide a more interpretable assessment. The prediction population should also be defined at the intended decision time. At the index PCI, subsequent survival is unknown. Mortality follow-up was available, yet patients who died within one year without TLR were excluded. These observed deaths preclude a later TLR and are competing events rather than missing outcomes. Removing them makes inclusion among non-cases depend on a post-baseline event. The fitted models therefore describe a competing-death-selected analysis population, not the one-year cumulative incidence among all patients eligible at PCI. Prediction-model validation in the presence of competing events should align the evaluated risk with the baseline target population.4 The publication does not report the number, timing, or covariate distribution of excluded deaths. The direction and magnitude of any resulting change in absolute risk therefore remain unknown. The investigators could retain all baseline-eligible patients and report death before TLR explicitly. They could then evaluate the fixed-horizon binary outcome with deaths retained as non-events, or estimate cumulative incidence when event timing is used. Either approach would show whether the model applies to patients for whom post-PCI follow-up decisions are actually made. The existing data permit a focused response. The authors could repeat performance assessment in the full baseline cohort and provide unrounded Brier benchmarks, calibration intercepts, slopes, and curves. If model-guided follow-up remains an intended use, they should link a prespecified risk threshold to a defined surveillance action and compare net benefit with follow-all and follow-none strategies.2,5 Until then, the models are better treated as developmental descriptions of predictor associations within the reported analysis population. Two questions would materially resolve the interpretation. Do predictions retaining competing deaths preserve the reported performance and calibration? Does any clinically justified threshold improve decisions? If not, narrowing the accuracy, target-population, and clinical-use claims would align the conclusions with the evidence currently reported. Funding Statement The author received no specific funding for this work. Author Contributions Jiaqin Liu: Conceptualization, Methodology, Formal analysis, Writing - original draft, Writing - review and editing. Disclosure The author declares no competing interests in this communication. References