Persistent postoperative anaemia and 1‐year mortality: re‐examining time origin and sample selection

We read with interest the study by Choi et al. [1], which examined the association between persistent postoperative anaemia and 1-year mortality after valvular heart surgery. By focusing on haemoglobin levels measured 2 months postoperatively, the authors question the prognostic significance of delayed haemoglobin recovery. We would like to offer our observations on this work that may have implications for the internal validity and generalisability of the results. There appears to be a misalignment between the exposure definition and the follow-up periods. Persistent anaemia was defined based on haemoglobin values measured 2 months after surgery, yet patients were followed up for mortality from the date of surgery. Patients who died within the first 2 months or lacked haemoglobin data at that time point were not included. This approach introduces a period of guaranteed survival between surgery and exposure ascertainment. Without aligning the timescale accordingly, such design carries the risk of immortal time bias [2]. Furthermore, restricting the analysis to patients who survived to and completed follow-up testing at 2 months may introduce survivor selection bias, limiting the generalisability of the results. We believe this issue could be addressed by adopting a formal landmark analysis, whereby only patients alive at 2 months with haemoglobin data are included and follow-up begins from that point forward [3]. Comparing the baseline characteristics and early postoperative outcomes between included and excluded patients would help evaluate the direction and magnitude of potential selection bias. We are concerned about the risk of overfitting in the multivariable model given the relatively small number of events. Only 81 deaths occurred in the cohort, yet the final model was derived using a combination of univariate screening, Akaike information criterion-based stepwise selection, least absolute shrinkage and selection operator and clinical judgement. While the number of covariates retained in the final model may appear modest, the complexity of the variable selection procedure increases the effective degrees of freedom and introduces a considerable risk of selection bias [4]. Moreover, model performance metrics such as the C-index, when evaluated on the same dataset used for variable selection, are likely to be biased optimistically. We encourage the authors to provide internally validated performance estimates (e.g. via bootstrapping or cross-validation) and, where possible, to report the events-per-variable ratio to contextualise model complexity. Simplifying the model by focusing on a small set of prespecified, clinically grounded predictors may further enhance its robustness and replicability. We commend the authors for highlighting the prognostic relevance of persistent postoperative anaemia. We believe that addressing the concerns above would improve the clarity and validity of the findings and provide stronger support for their potential clinical implications.