Comment on: “Lumbar Spine Endplate Sclerosis is a Protective Factor for Cage Subsidence in Minimally Invasive Transforaminal Lumbar Interbody Fusion”

Abstract

We read with great interest the study by Liao et al, 1 who report that lower endplate Hounsfield unit (HU) values and obesity independently predicted cage subsidence after minimally invasive transforaminal lumbar interbody fusion (MI-TLIF), and that an endplate HU cutoff of 221 stratified subsidence risk among patients with L1 HU < 117 (odds ratio 4.444, 95% CI 1.703-11.595). This is a clinically useful contribution, since endplate HU is readily obtainable from routine preoperative CT without additional radiation or cost. We would, however, welcome the authors’ comments on two aspects of the analysis that may affect how confidently the proposed threshold can be applied in practice. First, the HU cutoff of 221 was derived by maximizing the Youden index and then tested for association with subsidence within the same 213-endplate subgroup (L1 HU < 117) from which it was generated. Because both steps used the same dataset, the reported odds ratio is susceptible to optimism bias, and the true discriminative performance of this threshold in an independent cohort is likely to be more modest than observed here. 2 Did the authors perform any internal validation, such as bootstrap resampling or split-sample testing, to quantify the extent of this optimism before recommending 221 as an actionable cutoff, and do they plan to test this threshold in an external cohort or a different scanner/protocol before it is adopted more widely? Second, the primary analysis was conducted at the endplate level (464 endplates from 169 patients, with 37.9% of patients contributing two operated levels), yet obesity, bone quality, and other systemic patient-level factors would be shared by both endplates within the same patient. Treating endplates as independent observations when a substantial proportion are correlated within individuals can understate the true standard errors and narrow the confidence intervals around the reported odds ratios, particularly for endplate HU and obesity. Analytic approaches that explicitly accommodate clustered, non-independent observations, such as generalized estimating equations or mixed-effects models with a patient-level random effect, are well established for this type of data structure. 3 Could the authors clarify whether clustering by patient was accounted for in the multivariate logistic regression