Letter to ‘Risk Prediction Models for Frailty in Adult Maintenance Haemodialysis Patients: A Systematic Review and Methodological Appraisal’

Abstract

We read with great interest the recent article ‘Risk Prediction Models for Frailty in Adult Maintenance Hemodialysis Patients: A Systematic Review and Methodological Appraisal’ by Zhang et al. (2025), an important contribution to the field of predictive modelling in nephrology. The authors adeptly summarised methodological challenges and highlighted the critical gaps in the development and validation of frailty prediction models. However, through a deeper reflection, we would like to revisit and expand upon certain aspects of their findings, particularly regarding the methodological rigour, clinical utility, and future directions for research and implementation.

It is commendable that the reviewers placed a strong emphasis on the need for external validation to assess generalisability across diverse healthcare settings. However, the predominance of single-centre studies within the dataset (13 of 15 studies) raises significant concerns about selection bias and the representativeness of the included patient populations. For instance, populations from rural or resource-limited areas, where a significant proportion of maintenance haemodialysis patients reside, were underexplored. As frailty predictors, such as nutritional deficits or inflammatory markers, may vary based on socioeconomic disparities, how might predictive models account for this significant variability? Could multi-centre cohort studies more effectively balance these disparities, ensuring the development of equitable models?

Furthermore, a noticeable limitation was the underrepresentation of racial and ethnic minorities. As frailty phenotypes and biological markers can differ substantially across populations (Wong et al. 2021), a systematic effort to integrate globally representative cohorts is essential. This highlights the need for collaborative international research using standardised protocols.

The authors rightly stressed the inappropriate reliance on univariate analysis for variable selection in 9 of the included studies, a methodology deeply prone to omitting critical interaction effects and multicollinearity among predictors. However, our concern extends beyond this isolated critique. While the recommendation to incorporate advanced techniques, such as Least Absolute Shrinkage and Selection Operator (LASSO) regression, is warranted, we question the implications of overly complex models in clinical practice. Would the implementation of dynamic variables (as suggested, e.g., trends of albumin or interleukin-6) paradoxically lead to increased data burden, particularly in resource-constrained settings? How should researchers balance complexity with feasibility, ensuring utility beyond academic contexts?

The discussion could further benefit from assessing the role of explainability in machine learning models. While state-of-the-art methods, such as random forests or neural networks, have been employed in other specialisations (Debray et al. 2023), they often lack interpretability, posing challenges when applied in clinical decision-making for frail maintenance haemodialysis patients, where care must remain patient-centered. Clear communication between clinicians, patients and caregivers is crucial to ensuring guideline adherence.

A recurring methodological flaw identified was the frequent conversion of continuous variables into binary formats, observed in over 40% of the included models. This statistical simplification, while intuitive, leads to information loss and reduced predictive power (Lipkovich et al. 2017). In maintenance haemodialysis populations, clinical biomarkers, such as serum albumin or N-terminal pro-B-type natriuretic peptide (NT-proBNP), exhibit nuanced and non-linear relationships with frailty. Simplifying such data might mask critical threshold effects or interaction terms. Why then is this approach still so prevalent, despite its pitfalls? Does this reflect a persistent lack of statistical expertise among healthcare researchers?

Parallel to this issue, inadequate handling of missing data was highlighted, with a preference for complete-case analysis-rather than multiple imputation techniques-potentially introducing bias. Given these challenges, should future predictive modelling standards mandate the incorporation of imputation protocols? How might this standardisation further improve reproducibility and application?

The authors offered an extensive critique of the included studies' performance evaluation metrics, particularly noting reliance on the Hosmer-Lemeshow goodness-of-fit test in 11 studies. However, calibration metrics alone cannot always capture clinical relevance, particularly when discriminatory ability remains unquantified. Here, we align with the call for the integration of real-world validation, emphasising metrics such as Net Benefit or Clinical Utility functions (Moons et al. 2014). Does this create an opportunity to standardise such metrics within nephrology-specific TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) guidelines?

Moreover, none of the studies evaluated cost-effectiveness, an important consideration given that maintenance haemodialysis patients often face substantial financial burdens due to chronic treatment. Transparent thresholds for actionable predictions (i.e., optimal sensitivity/specificity trade-offs) must accompany such economic analyses to facilitate adoption.

While the article underscores the theoretical importance of predictive modelling in nephrology, clinical integration remains a significant hurdle. To ensure that these models inform rather than merely supplement decision-making, future approaches require alignment with treatment strategies. For example, models identifying frail patients at high risk of hospitalisation could prioritise comprehensive nutritional support, physical therapy interventions or deprescription trials to address polypharmacy. How can predictive models be integrated with electronic health records (EHRs) or mobile decision-support applications? Cross-specialty learning from cardiovascular disease or diabetes management, both of which leverage EHR-linked real-time decision tools, could offer valuable insights.

Additionally, we echo the suggestion for incorporating dynamic predictors. Biological markers such as granulocyte-colony stimulating factor or growth differentiation factor-15 (GDF-15), cited as emerging targets, warrant further exploration. At the same time, efforts must consider their availability and cost-effectiveness in resource-limited settings, lest these models inadvertently widen existing healthcare inequalities.

In summary, while the reviewed article illuminates critical weaknesses within existing frailty prediction models for maintenance haemodialysis patients, this letter highlights the need for pragmatic and globally focused refinements. Addressing biases in population selection, ensuring transparency in variable handling, and expanding performance evaluation metrics are crucial prerequisites for actionable models. Parallel investments in implementation science will likewise be necessary to ensure their seamless integration into care pathways, allowing them to meaningfully impact patient outcomes. By fostering interdisciplinary collaboration across clinicians, data scientists and public health experts, the predictive models of tomorrow can achieve both sophistication and usability. We thank the authors once again for their valuable contribution and hope this dialogue will further enrich future research within this rapidly evolving field.

The authors have nothing to report.

The authors declare no conflicts of interest.

The authors have nothing to report.