We read with great interest the recent article by Duan and colleagues, which presents an explainable machine learning approach to predict 10-year incident cognitive impairment among individuals with early-stage cardiovascular–kidney–metabolic (CKM) syndrome.1 The study is notable for translating a contemporary cardiometabolic framework into an implementable clinical tool by leveraging routinely collected variables, benchmarking multiple algorithms, and providing transparent individualised explanations through SHAP alongside an accessible web-based calculator. This combination of feasibility, predictive performance, and interpretability offers a pragmatic template for integrating “brain health” prevention into cardiometabolic care. While the authors' contribution is substantial, several methodological considerations that are not fully explored may influence how the reported long-term risk should be interpreted. First, prediction over a decade in an ageing cohort is intrinsically shaped by informative attrition and competing events, particularly mortality. When participants at higher cardiometabolic risk are more likely to die or to be lost before outcome ascertainment, the model may preferentially learn the probability of cognitive impairment among those who remain under observation, rather than the risk in the baseline population. This distinction matters clinically because it can lead to optimistic estimates of absolute risk and attenuated calibration in those with the greatest vulnerability.2, 3 Future work would be strengthened by explicitly accounting for competing risks and differential follow-up so that the target of prediction aligns more closely with real-world clinical populations and decision contexts. Second, the model relies primarily on baseline measurements, yet cardiometabolic risk and cognitive vulnerability are dynamic and strongly influenced by subsequent risk-factor control and treatment initiation. A baseline-only approach may conflate initial severity with long-term prognosis, thereby limiting the extent to which the model can support truly modifiable prevention strategies. Incorporating longitudinal information, such as changes in blood pressure, glycemic status, lipid levels, and therapeutic exposure over follow-up, would help distinguish persistent risk from risk that is mitigated by effective management and would likely increase both clinical interpretability and actionability of the predicted probabilities.4 Third, the operational definition of incident cognitive impairment, although practical for large cohort studies, may interact with educational and sociocultural factors that influence test performance independent of underlying neurodegenerative processes.5 If education contributes to outcome classification through measurement properties of cognitive testing, its prominence as a predictor may partly reflect criterion contamination rather than etiologic importance, which can subtly reshape the interpretation of feature attribution and risk pathways. Strengthening outcome modelling by reducing measurement dependence on educational context and demonstrating stable calibration across educational strata would further support both the robustness and the clinical fairness of the tool. In conclusion, Duan and colleagues provide an important and timely contribution by integrating early CKM profiles with explainable machine learning to estimate long-term cognitive risk using readily available clinical variables. Addressing the implications of competing events and informative attrition, moving towards dynamic prediction that reflects treatment-responsive trajectories, and reinforcing outcome definitions against educational measurement effects would further enhance the validity and translational value of this promising framework. All authors contributed to the study conception and design. Study design: Jiaxuan Li, Xin Miao, Zongbao Li; Writing—original draft: Jiaxuan Li, Jie Yang; Writing—review and editing: Jiaxuan Li, Xin Miao, Jie Yang; Supervision: Jie Yang. None. None. The authors declare no conflicts of interest. The peer review history for this article is available at https://www.webofscience.com/api/gateway/wos/peer-review/10.1111/dom.70437. Data sharing is not applicable to this article as no new data were created or analysed in this study.
Li et al. (Mon,) studied this question.