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

Abstract

We are writing in response to the article titled ‘Risk Prediction Models for Frailty in Adult Maintenance Hemodialysis Patients: A Systematic Review and Methodological Appraisal’ (Zhang et al. 2025). The authors should be commended for their effort in synthesising evidence on frailty risk prediction models in maintenance haemodialysis (MHD) patients, a population notoriously vulnerable to adverse outcomes. However, we wish to highlight certain aspects of the review that merit further discussion.

First, although the authors meticulously assessed methodological robustness using tools, such as TRIPOD, one cannot help but question whether prediction model development in the reviewed studies adequately reflects the complexities of frailty in MHD patients. Frailty in MHD is undeniably multifactorial, influenced by age, dialysis-related factors, inflammation, nutrition and psychosocial status. However, the inclusion of univariate analysis for variable selection in nine studies raises concerns about oversimplified model construction (Lipkovich et al. 2017). Should predictive modelling efforts not prioritise multivariate strategies, such as LASSO regression, to capture nuanced interactions and address multicollinearity (Efthimiou et al. 2024)? This would align with the inherent complexity of frailty and may improve generalisability across diverse MHD populations. Furthermore, binary transformation of continuous variables, as noted in multiple studies (Zhang et al. 2025), significantly reduces data informativeness. For example, converting ‘serum albumin’ into a dichotomous variable ignores clinically significant gradients that can stratify patients' frailty risk more precisely.

Second, although external validation was rightfully emphasised as critical, its scarcity in practice is worrying. The single externally validated model demonstrated an AUC decline of 15% when applied to independent cohorts (Zhong et al. 2025). This highlights not only a lack of generalisability but also potential overfitting during model development, which is further compounded by small sample sizes across studies. Given the need for predictions that transcend centre-specific practices, wouldn't a collaborative, multi-centre initiative with larger, ethnically diverse cohorts improve the validity and adoption of prediction tools? Such efforts could also ensure that region-specific practices, such as dialysis modalities or nutritional interventions, are reflected in frailty risk models.

Third, from a clinical standpoint, the risk models lack actionable integration into routine practice. Despite proposing dynamic tools like wearable-derived metrics, their visibility within the reviewed studies was limited. Considering the rapid adoption of wearable technologies in nephrology, how might future models incorporate real-time patient data—for example, step counts or heart rate variability—to enhance FRAIL predictions dynamically during MHD sessions? Additionally, although several studies assessed traditional biomarkers like IL-6 and GDF-15, their inclusion feasibility remains challenging for resource-limited clinics. Practical questions, therefore, arise: Should prediction models prioritise inexpensive, widely available biomarkers, such as serum calcium or haemoglobin levels to improve scalability within resource-strained healthcare environments?

Fourth, the ambiguity surrounding frailty measurement tools warrants closer scrutiny. Over 60% of the reviewed studies employed the FRAIL scale, but its subjectivity introduces significant inter-assessor variability (Kennard et al. 2023). With multiple validated frailty metrics (e.g., Fried Phenotype, Clinical Frailty Scale) available, would a consensus guideline selecting the most predictive and feasible assessment tool for MHD patients not strengthen both research and practice? Furthermore, standardised training protocols for assessors could minimise the bias introduced by subjective interpretations, a concern unaddressed in four of the included studies.

Fifth, the review highlights poor reporting transparency among the studies assessed. Specifically, adherence to TRIPOD guidelines was limited to two studies, with a prevalent omission of model equations and missing data flow diagrams. Such omissions significantly hinder reproducibility and transparency.

In summary, the review robustly outlines current gaps in frailty risk prediction models for MHD patients. In view of this, there are several points that can be considered as future research directions: (1) adoption of multivariate modelling strategies that embrace frailty's multifactorial nature; (2) collaborative research efforts focusing on multi-centre, ethnically diverse cohorts to maximise external validity; (3) integration of dynamic wearable technologies into frailty scoring models for real-time assessments; (4) prioritisation of accessible, cost-effective biomarkers to ensure scalability; and (5) standardising frailty assessment tools and improving reporting practices through adherence to international guidelines like TRIPOD (Efthimiou et al. 2024). These efforts would provide clinicians with more reliable, generalisable and actionable frailty prediction tools.