Limitations of applying the COX proportional hazards model to glioma studies

Dear Editor,

We read with great interest the recently reported work by Geng et al. [1], entitled “SIRPB1 regulates inflammatory factor expression in the glioma microenvironment via SYK: functional and bioinformatics insights.” In this study, the authors conducted both univariate and multivariate Cox regression analyses based on 670 glioma samples from the cancer genome atlas (TCGA) database. According to the results of multivariate analysis, isocitrate dehydrogenase (IDH) status, primary therapy outcome, age, and signal regulatory protein beta 1 (SIRPB1) expression were identified as independent overall survival (OS) prognostic factors. Following that, making use of SIRPB1 expression along with additional clinicopathological variables, they constructed a prognostic model for 1-, 3-, and 5-year OS that showed a high degree of prediction accuracy.

However, we observed certain deviations in the authors’ application of the Cox proportional hazards (CoxPH) model, which were not addressed by the authors. Both interval-censoring and right-censoring events are possible outcomes of the established standards [2]. Glioma prognostic events, diagnosed via medical records, may lead.

to interval censoring if they transpire between follow-up visits and to right censoring if diagnosed between the conclusion of follow-up and the data analysis phase. The CoxPH model primarily focuses on handling right-censored data. Conversely, the accelerated failure time (AFT) model is typically preferred for scenarios encompassing a variety of censored data types [3]. The “icenReg” and “survival” packages can be used to estimate event timings and fit and analyze mixed censored data.

From the perspective of model construction, the CoxPH model necessitates the assumption of proportional hazards (PH), implying that the impacts of covariates remain constant across time. However, a number of reasons may contribute to the frequent emergence of nonproportionality of hazards in practice. The authors ought to employ Schoenfeld residuals or other techniques to assess the PH assumption in relation to the covariate-outcome connection [4]. If the residuals demonstrate a consistent pattern of change over time, it indicates that the covariate’s effect may vary with time. In cases where the PH assumption is violated, the authors should utilize a Cox model incorporating time-varying effects or an accelerated failure time model rather than the conventional CoxPH model [3, 5].

Similarly, we employed data featuring clinical parameters equivalent to those adopted by Geng et al. [1] in their investigation of glioma patients from the TCGA cohort, performing both univariate and multivariate Cox regression analyses, followed by an assessment of the PH assumption underlying the multivariate Cox regression model. Subsequently, based on the results of the multivariate Cox regression, we designed a nomogram model for 1-year, 3-year, and 5-year OS (Fig. 1A) and further generated corresponding calibration curves (Fig. 1B). The findings indicated that the IDH mutation status, SIRPB1 expression level, and the global test failed to meet the PH hypothesis (Table 1), potentially compromising the statistical credibility and precision of the OS prediction model proposed by Geng et al. [1].

Fig. 1.

Construction and evaluation of the nomogram. (A) Nomogram for 1-, 3-, and 5-year OS based on IDH status, primary therapy outcome, age, and SIRPB1 expression. (B) Calibration curves showed the concordance between predicted and observed 1-, 3-, and 5-year OS

Table 1.

PH assumption test to multivariate Cox regression

Characteristic	Chi-square value	Degree of freedom	P-value
IDH status	13.647	1	0.0002
Primary therapy outcome	0.962	1	0.3267
Age	3.379	2	0.0660
SIRPB1	4.299	1	0.0381
Global	17.697	4	0.0014

If the p-value in the global test > 0.05, it indicates that the multivariate Cox regression adheres to the PH assumption

In summary, we deem it essential to conduct a reassessment, taking into account the potential ramifications of censoring events and the proportional hazards assumption. Despite our concerted efforts to emphasize the prerequisites for adopting multi-factor Cox regression and creating nomograms for researchers, a substantial number of publications persist in neglecting the PH assumption in their studies. This highlights the fact that rigorous model validation remains an overlooked yet crucial aspect in scientific research, warranting heightened attention and priority.

Abbreviations

TCGA

The Cancer Genome Atlas

IDH

Isocitrate Dehydrogenase

SIRPB1

Signal Regulatory Protein Beta 1

OS

Overall Survival

PH

Proportional Hazards

Author contributions

JX, YC, CX, QY, and RL wrote the draft. LC and WM rewrote it. All authors read and approved the final manuscript. Jihao Xue, Yitian Chen, and Cheng Xue contributed equally to this study.

Funding

This research was funded by the National Natural Science Foundation project (No. 82072780 and 82372825), the Science and Technology Department of Sichuan Provincee (No. 2022YFS0630). This research was further supported by the Natural Science Foundation of Southwest Medical University, Grant (No. 2023QN008).

Data availability

The dataset analyzed for this study can be found in the https://xenabrowser.net/datapages/.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have approved the manuscript for publication.

Competing interests

Not applicable.