Assessing time-trend bias in glioblastoma prognosis over 2 decades of clinical trials

Abstract

Background

The time-trend bias represents a potential limitation in the use of external controls in glioblastoma (GBM) trials. In this study, we assessed whether outcomes for newly diagnosed GBM (ndGBM) patients treated with the standard Stupp protocol in clinical trials have changed over the past 2 decades.

Methods

We retrieved individual patient-survival pseudo-data from Stupp-protocol arms reported in trials published over the last 20 years. Survival distributions were approximated using Weibull distributions, and an Accelerated Failure Time model was used to evaluate any potential time-trend by correcting for identified key prognostic factors.

Results

MGMT methylation status and Karnofsky Performance Status emerged as the main determinants of survival differences among clinical trials. Both in a multivariable regression that included all candidate prognostic factors and after adjustment for the main determinants, the publication year showed no impact on the outcome of the Stupp-protocol control arms. The performance of the model was validated using 3 independent Phase III cohorts, providing additional evidence for the absence of time‐trend bias.

Conclusions

No evidence of time‐trend bias was observed in Phase III GBM trials over the past 2 decades once major prognostic factors were accounted for.

Key Points.

• No time‐trend bias was observed in clinical trials for newly diagnosed glioblastoma over the past 20 years.

• The absence of time-trend bias was validated using 3 independent cohorts derived from Phase III randomized controlled trials.

Importance of the Study.

The use of external controls in clinical trials in neuro-oncology is potentially affected by methodological limitations. Among them, the time-trend bias is based on the assumption that patient prognosis has improved over time. In this study, we analyzed the control arms following the Stupp protocol in all Phase III clinical trials conducted over nearly 20 years of neuro-oncology clinical research. Leveraging the availability of individual patient-survival pseudo-data, we investigated whether the time of publication influenced survival probability distributions. Our analysis shows that, once adjusted for major prognostic factors, the prognosis of patients with newly diagnosed glioblastoma has not changed over this extended period. While this work mitigates concerns about the use of external controls from previous clinical trials, it also highlights a concerning lack of progress in improving glioblastoma patient outcomes within the consistent framework of Phase III clinical trials.

With a 5-year survival rate of only 6.8%,¹ glioblastoma (GBM) remains one of the most aggressive and lethal cancers. Despite significant advances in brain cancer research over the past 2 decades,² no new treatment—aside from tumor-treating fields (TTF)³—has demonstrated efficacy in a randomized Phase III clinical trial.⁴ Given the urgent need for new therapeutic options, the slow pace of drug development, and the considerable ethical and logistical challenges in enrolling control arms in GBM clinical trials,⁵ the use of external controls during the clinical development pipeline would be of great interest. However, the use of this carries several methodological challenges, which were comprehensively addressed in a recent review.⁶ Among the potential biases discussed, the authors highlighted time-trend bias, which refers to the improvement in overall survival (OS) over time in patients treated with the same standard of care. This effect appears plausible in neuro-oncology due to progress in radiotherapy and neurosurgery,⁷ the improvements in supportive care and early detection of iatrogenic side effects. However, the current evidence on this topic is conflicting. Indeed, while this widely accepted principle has been supported by retrospective data,⁶ a recent study by Rahman et al. reached the opposite conclusion by analyzing the median overall survival (mOS) of trials conducted over the past years.⁸

In this study, to address the issue of time-related bias, we leveraged a comprehensive database of individual patient-survival pseudo-data from the Stupp-protocol arms of all Phase III clinical trials conducted in newly diagnosed GBM (ndGBM) over the past 2 decades.⁹ This database enables modeling of predicted survival probabilities and assessment of the impact of covariates on patient outcome within a homogeneous setting. The impact of publication year on survival variability across trials was therefore evaluated in light of the known prognostic factors for GBM, finding no evidence of a covariate-adjusted time-trend bias in GBM prognosis over the past 2 decades.

Materials and Methods

Data Collection

Data for the analysis were retrieved from our previous publication.⁹ In that study, we conducted a systematic review of the literature to identify all Phase III randomized controlled trials (RCTs) on ndGBM that included the Stupp protocol as one of the treatment arms, published between January 2003 and July 2023. From the 13 included studies, Kaplan–Meier curves were extracted and digitized to generate individual patient-survival pseudo-data. In addition, we collected demographic and baseline clinical characteristics of the arms treated with the Stupp protocol. For patient age, when the mean was not explicitly reported, we estimated it using the median and interquartile range, or the reported range, according to the method described by Wan et al.¹⁰ After excluding 2 trials that included only patients expressing EGFRvIII,^11,12 11 studies were retained for our analysis^3,13–22 (see Table 1).

Table 1.

Baseline Characteristics of the Stupp Protocol Arms in the Included Studies

Author, year	N	Sex (male%)	Age (mean)	KPS ≥ 90 (%)	Methylated MGMT (%)	Gross total resection (%)	Excluded early progressions
Stupp et al., 200513	287	64.5	50.2a	39.4	43.4	39.4	No
Gilbert et al., 201314	411	58.2	55.0a	66.4	32.4	56.0	Yes
Chinot et al., 201415	463	64.4	52.3a	69.7	33.7	42.3	Yes
Gilbert et al., 201416	309	62.8	53.8a	61.9	28.4	60.3	Yes
Stupp et al., 201417	273	52.4	60.7a	55.5	100.0	50.6	Yes
Westphal et al., 201518	71	63.4	53.0	NA	33.3	42.3	No
Kong et al., 201719	89	57.3	55.8	NA	NA	53.9	Yes
Stupp et al., 20173	229	68.6	53.3a	66.8	44.8	53.7	Yes
Herrlinger et al., 201920	63	47.6	58.3a	77.8	100.0	63.5	No
Lim et al., 202221	358	55.0	54.8a	70.3	98.0	55.9	No
Omuro et al., 2023222	280	62.5	54.0a	75.4	0.0	51.4	No

Abbreviations: KPS, Karnofsky Performance Status; N, number of patients.

^aData were estimated using the median and interquartile range (or the overall range) reported in the studies.

Statistical Considerations

The reconstructed survival pseudo-data were tested pairwise for the proportional hazards (PH) assumption using Schoenfeld residuals tests.²³ The test results are reported as P-values in Supplementary Table 1, where P < .05 indicates a violation of the PH assumption.

Due to multiple violations of the PH assumption and the limited number of observations (11 trials), which restricts the feasibility of fitting a non-PH model or incorporating time-dependent interactions, we opted for an Accelerated Failure Time model to explore the effect of covariates on patient survival.²⁴ This approach allows us to directly model survival time instead of hazards.

The empirical time-dependent variation of the hazard function across studies was evaluated by computing empirical hazard rates using reconstructed survival data. The hazard rate at a given time was defined as:

$${\displaystyle \begin{array}{l}{\displaystyle h\left(t\right)=\frac{d\left(t\right)}{n\left(t\right)}}\end{array}}$$

where d(t) represents the number of events occurring at that time and n(t) is the number of patients still at risk at that time. Nine out of the 11 included populations, representing 79% of the total sample, exhibited a monotonically increasing hazard function. A simplified graphical representation is provided in Supplementary Figure 1, where the hazard functions were smoothed using locally weighted scatterplot smoothing with a fractional parameter of 0.6 to reduce noise due to small sample sizes at later time points. Based on this trend, a parametric Weibull distribution was assumed to model the survival curves.²⁵

Estimation of Weibull Parameters

For each trial’s survival curve, reconstructed from individual patient-survival pseudo-data, we applied the WeibullFitter class from the lifelines Python library. Specifically, we fitted a 3-parameter Weibull model, constraining the location parameter to zero to ensure consistency and comparability across different datasets. The shape (k) and scale (λ) parameters were estimated via maximum-likelihood estimation and are reported in Table 2. The confidence interval (CI) for the Weibull shape parameter k was derived using the asymptotic normal approximation. The table also includes the corresponding mOS values as reconstructed from survival pseudo-data. As previously reported in Sferruzza et al.,⁹ these values differ from those in the original paper by less than 3%.

Table 2.

Summary of Median Overall Survival and Estimated Weibull Parameters

Author, year	mOS	Shape (k) (95% CI)	Scale (λ)	Location
Stupp et al., 2005	14.63	1.35 (1.20–1.49)	21.20	0.00
Gilbert et al., 2013	16.47	1.24 (1.23–1.35)	23.42	0.00
Chinot et al., 2014	16.63	1.61 (1.47–1.75)	22.28	0.00
Gilbert et al., 2014	16.27	1.42 (1.25–1.59)	21.99	0.00
Stupp et al., 2014	26.80	1.21 (1.04–1.39)	40.74	0.00
Westphal et al., 2015	19.10	1.72 (1.33–2.12)	26.08	0.00
Kong et al., 2017	16.53	1.36 (1.06–1.66)	26.97	0.00
Stupp et al., 2017	16.12	1.39 (1.24–1.55)	22.57	0.00
Herrlinger et al., 2019	31.41	1.34 (0.98–1.70)	50.41	0.00
Lim et al., 2022	31.94	1.72 (1.53–1.92)	38.46	0.00
Omuro et al., 2023	14.84	1.85 (1.65–2.05)	18.51	0.00

Abbreviations: CI, confidence interval; mOS, median overall survival.

The goodness of fit was evaluated graphically by comparing the reconstructed Kaplan–Meier curves with the fitted Weibull distributions (Supplementary Figure 2, left panels). Additionally, we assessed model adequacy by computing Cox–Snell residuals and plotting the Nelson–Aalen cumulative hazard of these residuals against the 1:1 reference line.²⁶ The closer the residual plot aligns with this diagonal, the better the Weibull model represents the observed survival data (Supplementary Figure 2, right panels).

Analyze the Impact of Predictors on Expected Survival

The parameters of the fitted Weibull distributions (Table 2) demonstrated relatively low variability in the shape parameter (k) across the studies. Furthermore, the reconstructed mOS, which in this model is determined by the scale λ and shape k parameters according to the formula,

$${\displaystyle \begin{array}{l}{\displaystyle mOS=\lambda \ast {\left(\log \left(\text{2}\right)\right)}^{\frac{1}{k}}}\end{array}}$$

was found to be strongly correlated with the scale parameter λ (Pearson correlation coefficient = 0.94, P < .001). These findings suggest that the influence of covariates on patient survival can be approximated by assessing their effect on the scale parameter λ. Thus, we applied linear regression models to investigate the impact of covariates on λ, following a natural log transformation.

To assess the effect of time-trend on patient-survival outcomes, we first performed a linear regression analysis to predict log(λ) as a function of the indexed year of publication. We then performed a second analysis by running a multivariate regression model, adjusting the indexed publication year for all known prognostic factors (KPS ≥ 90, MGMT methylation status, percentage of gross total resection, mean age, sex, and exclusion of early progressions). Subsequently, to strengthen the main analysis and reduce coefficient instability due to the number of covariates relative to the number of included trials, we ran a second multivariate regression adjusting the effect of publication year only for the most influential covariates. To identify the most influential covariates, we used a stepwise backward elimination procedure. At each iteration, we removed the least significant predictor based on P-values (α = 0.05) and refitted the model. This process continued until all remaining predictors were statistically significant. Finally, we performed an analysis of variance (ANOVA) decomposition on the final model to quantify the net percentage of variance explained by each predictor.

To evaluate the robustness of the proposed analysis, we conducted a leave-one-out sensitivity analysis on the final model. For each study in our dataset, we excluded all observations from that study and refitted the multiple linear regression model that included the publication year and the most influent covariates.

We conducted 3 additional analyses by adjusting the effect of the indexed publication year for mean age, percentage of gross total resection, and the combination of these 2 variables, selected because of their major prognostic relevance in GBM patients.^27–32

Furthermore, since one of the included studies²² excluded patients with IDH1 and IDH2 mutations, we conducted a dedicated sensitivity analysis to rule out potential bias by repeating both the univariate and multivariate analyses after excluding this study.

Correcting for the Potential Confounding Effect of the TTF

TTF was approved by the FDA in 2011 and has since been increasingly adopted in clinical practice. Therefore, we considered whether this variable may have influenced our results. Although TTF use was not included in any of the protocols of the included studies, it remains possible that some patients received TTF after relapse. Given the proven efficacy of TTF in RCT,³ this may represent a potential source of time-trend bias in OS. To account for this potential confounder, we generated a virtual variable to approximate the proportion of patients who may have received TTF as a subsequent-line treatment. We averaged real-world data on TTF use at recurrence from two academic neuro-oncology centers (New York, USA, and Heidelberg, Germany), as reported by Lassman et al.³³ The authors reported that approximately 4% of recurrent GBM patients received TTFs between 2011 and 2014, increasing to 16% between 2015 and 2019. Assuming a constant rate of increase for simplicity, we modeled a linear growth of 2.67% per year starting in 2011, which closely mirrors the observed trend in real-world data.³³ This virtual variable was then included in the analysis to assess its potential confounding effect in the evaluation of the time-trend bias.

Model Validation

To validate our results using external real-world data, we identified 3 Phase III clinical trials^34–36 that were not included in our study because they were published after the end of the systematic review period. From these trials, we extracted baseline data from their Stupp-protocol control arms, encompassing 686 patients in total, to test whether patient survival can be predicted without reference to the study’s publication year but using only the main predictors that emerged from the stepwise backward elimination procedure. The proportion of participants with KPS ≥ 90 was not reported in the manuscript by Sarkaria et al.,³⁴ but was kindly shared by the corresponding author on request (42.86%). To propagate the uncertainty of the regression model to the predicted OS, we used a Monte Carlo procedure³⁷ as follows:

1. Sampling the regression coefficients to estimate the scale parameter λ: At each iteration, we drew an entire coefficient vector from the multivariate normal distribution defined by the point estimates and their full variance–covariance matrix, thereby preserving the empirical correlation among the regression parameters.

2. Sampling the Weibull shape parameter k : The shape estimates for the Phase III trials included in our study (see Table 2) were fitted to a log-normal distribution by maximum likelihood; one value of k was drawn from this distribution at every iteration to reflect interstudy heterogeneity.

For every simulated pair of scale λ and shape k, we calculated mOS with the formula described earlier. The cycle was repeated 50 000 times. We summarized the simulated mOS distribution by reporting its median value (point estimate) and the corresponding 95% CI. The overlap between this simulated CI and the CI reported in the literature served as the basis for our hypothesis test.

Statistical Analysis

All data analyses and figure generation were conducted using Python 3.9 within a Jupyter Notebook environment and R 4.4. All codes will be made available upon reasonable request.

Results

The Impact of Baseline Predictors on Expected Survival

The backward linear regression analysis identified MGMT methylation status and KPS as the main predictors (see Table 3) as the primary predictors of survival variability across the included trials.

Table 3.

Stepwise Backward Linear Regression for Predicting log(λ)

Variable	Coeff.	95% CI	Std. error	t-Value	P-value	Explained variance (%)
Intercept	2.38	(1.95–2.8)	0.174	13.685	<.001	–
KPS ≥ 90	0.007	(0.001–0.013)	0.003	2.575	.04	4.68
MGMT (methylated)	0.009	(0.007–0.011)	0.001	11.368	<.001	91.09

Abbreviations: ANOVA, analysis of variance; CI, confidence interval; KPS, Karnofsky Performance Status.

A stepwise backward elimination procedure was performed to identify the most parsimonious set of predictors for log(λ). Starting with all candidate predictors in the model, the predictor with the highest P-value exceeding .05 was iteratively removed. An ANOVA decomposition was then computed on the final model to quantify the net percentage of variance explained by each remaining predictor.

Model fit: adjusted R² = 0.94.

Overall significance: F = 68.20; P = < .001.

Notably, MGMT promoter methylation status appears to explain the majority of the variability in λ (91.1% of the explained variance, as reported in Table 3), and consequently, the OS variability observed in these Phase III clinical trials.

The Effect of Time-Trend Bias on Patient-Survival Outcomes

The indexed year of publication, which showed no significant effect in the univariate model (β = 0.04, P = .49; see Figure 1a), was found to be entirely noninfluential after adjusting for all the prognostic factors (β = −0.04, P = .48) and for the 2 main predictors MGMT methylation status and KPS (β = −0.01, P = .65, see Table 4).

Figure 1.

Analysis of the time-trend of the scale parameter λ: (a) Scatter plot showing the observed values of natural logarithm of the scale parameter λ (log(λ)) plotted against the indexed study year, with a linear regression line illustrating the trend over the years. (b) Adjusted predictions of log(λ) using values for performance status (KPS ≥ 90 = 39.4%) and proportion of patients with methylated MGMT promoter (43.4%) consistent with the 2005 EORTC study, plotted against the study timeline. This illustrates the negligible effect of time under controlled conditions.

Table 4.

Leave-One-Study-Out Analysis for the Effect of Indexed Year on log(λ), Adjusted for MGMT Status and KPS

Study removed	Coeff.	P-value
–	−0.01	.65
Stupp et al., 2005	−0.01	.61
Gilbert et al., 2013	−0.01	.81
Chinot et al., 2014	−0.02	.45
Gilbert et al., 2014	−0.01	.72
Stupp et al., 2014	−0.01	.69
Stupp et al., 2017	−0.01	.74
Herrlinger et al., 2019	−0.005	.79
Lim et al., 2022	0.003	.89
Omuro et al., 2023	−0.05	.11

Abbreviations: KPS, Karnofsky Performance Status.

Each row shows the coefficient (β) and P-value for indexed year from a linear model that also includes the MGMT status and KPS as covariates, after excluding data from one study at a time.

In Figure 1b, we provide an intuitive representation of this result by plotting the adjusted predictions of log(λ) for each study using the baseline predictor values from the 2005 EORTC study: KPS ≥ 90 = 39.4% and MGMT promoter methylation = 43.4%. This finding was further reinforced by the leave-one-out sensitivity analysis presented in Table 4, which demonstrated that the result was not biased by any single included study. Furthermore, even after adjusting the model for patient age, the percentage of gross total resection, the interaction between these 2 variables, and the estimated proportion of patients who received TTF as subsequent therapy, no time-trend in GBM prognosis was detected (β = 0.01, P = .82; β = 0.01, P = .87; β = 0.01, P = .84; and β = 0.07, P = .83, respectively).

To evaluate whether the exclusion of patients with IDH1 and IDH2 mutations in 1 study²² may have influenced the assessment of time-trend bias, we conducted a dedicated set of sensitivity analyses by excluding Omuro et al.²² The leave-one-out analysis reported in Table 4 shows that, after excluding this study, the effect of publication year appeared not to be influential when adjusted for MGMT promoter methylation and KPS ≥ 90 (β = −0.05, P = .11). Furthermore, excluding Omuro et al. 2023, the publication year was not significantly associated with patient prognosis in the univariate analysis (β = 0.10, P = .06), nor when adjusted for patient age (β = 0.06, P = .09), the percentage of patients who underwent gross total resection (β = 0.08, P = .18), or the combination of these 2 variables (β = 0.07, P = .09).

Model Validation

Using the model parameters reported in Table 3 and the estimated shape parameter k (Table 2), for Roth et al.,³⁵ we predicted from the baseline characteristics of the Stupp-protocol arm a mOS of 17.4 months (95% CI, 15.7–19.0), differing from the trial result by only 11 days (2% discrepancy; observed mOS: 17.0 months, 95% CI, 15.9–18.6). For this study, the agreement between the observed survival distribution (digitized from the Kaplan–Meier curve as described in Sferruzza et al.⁹) and the Weibull distribution predicted by our model is shown in Supplementary Figure 3. For Sarkaria et al.,³⁴ we estimated an mOS of 28.1 months (95% CI, 23.9–32.8), which not significantly overestimated the reported value of 24.8 months (90% CI, 22.6–27.7). Lastly, for Laprie et al., our model predicted a median survival of 19.7 months (95% CI, 17.9–21.5 ), while the study reported 22.6 months (95% CI, 18.9–25.4).

Discussion

The potential impact of time-trend bias, due to a drift in patients’ OS over time,⁶ on comparisons of GBM cohorts from different time periods remains a matter of debate. This issue holds particular relevance in neuro-oncology, as the presence of such bias represents a key limitation in leveraging the full pool of external control data available from past GBM trials.⁶ Indeed, while RCTs remain the gold standard for evaluating therapeutic efficacy, the use of external control arms has gained attention in neuro-oncology for their potential to accelerate drug development and approval.⁵ Although methodological strategies to address temporal drift have been proposed,³⁸ the discussion should begin by establishing whether a time-trend in GBM prognosis actually exists, and if so, to what extent it may limit the use of historical cohorts as external controls. Studies addressing this issue in recent years have reached conflicting conclusions. Specifically, Grossman et al. have proposed in 2010 the drift of patient OS over time as an explanation for the prognostic differences observed between the 2005 EORTC cohort and with NABTT cohort, in which patients were treated with radiotherapy and temozolomide only or 1 of 3 other medications.³⁹ However, the EORTC and the NABTT patients treated with radiotherapy and temozolomide only had similar survival with no ‘time-trend bias’ (their Table 2). In another study, Thomas-Joulié et al. reported a longer OS with the standard of care by comparing cohorts of patients diagnosed between 2005–2012 and 2013–2018.⁴⁰ However, due to the substantial amount and imbalance among the 2 cohorts of missing data for the key prognostic factor MGMT methylation status (reaching 77% missing data in the 2013–2018 cohort), along with the monocentric and retrospective design, the study’s conclusions are inherently limited by potential confounding factors. In contrast, a recent publication challenged the existence of a time-trend drift in GBM,⁸ showing that over the past 10 years, across a heterogeneous group of Phase II and III clinical trials, there was no evidence of a time-dependent drift in mOS.

In our study, we addressed the same issue by leveraging the availability of individual patient pseudo-data from eleven Phase III clinical trials conducted in ndGBM over the last 2 decades.⁹ By estimating their survival probability distribution, we demonstrated that the year of publication had no impact on the outcome of Stupp-protocol arms once corrected for the main prognostic factors of GBM patients. In contrast, patient outcomes among the included trials were predominantly influenced by MGMT promoter methylation status and baseline performance status. It is important to underline that this analysis was aimed at solely exploring the weight of the known prognostic factors in explaining the survival difference among the Phase III clinical trials and selecting the most relevant for a more robust correction of the impact of the publication year on patient survival. As such, these findings cannot be considered a comprehensive evaluation of their prognostic effect in GBM patients, as these data would not be appropriate for this purpose.

We also explored the possibility that the TTF, introduced in 2011, could act as a hidden variable influencing the evaluation of time-trend bias. By modeling a variable that captured the adoption of TTF as a subsequent line of therapy, based on real-world usage data,³³ we were able to exclude this potential confounder. This further supports the conclusion that no meaningful time-related shift in GBM patient prognosis has occurred over the past 2 decades.

To further test the robustness of our conclusions, we validated our model using 3 additional Phase III RCTs^34–36 that were not included in our original analysis, as they were published after the completion of the systematic review.⁹ All in all, the predictions we generated for these trials, obtained without including publication year as a covariate, show that the control-group outcomes do not significantly deviate from those estimated by a model built on data from the past two decades. In other words, this validation further supports the absence of any substantial time-related drift in the prognosis of GBM patients.

It is important to acknowledge the limitations of our study. The primary concern relates to potential unmeasured confounders present across the included studies. Specifically, we were unable to incorporate the potential prognostic impact of baseline corticosteroid use, which is known to influence GBM outcomes,⁴¹ or the use of bevacizumab as a subsequential therapeutic line. Furthermore, our analysis did not directly account for several key factors that have been proposed as plausible contributors to a time-trend bias in neuro-oncology, such as advancements in radiotherapy, neurosurgical techniques, diagnostics, and overall patient care. However, the absence of any time-trend drift in the context of Phase III clinical trials suggests that the combined effects of these factors have not had a measurable impact on patient prognosis within this setting.

A similar conclusion was recently reached by Rahman et al.,⁸ who evaluated Phase II and III clinical trials conducted over a 10-year period. Their analysis demonstrated no change in mOS over time once adjustments were made for the prognostic factors. The fact that 2 independent studies, using different methodologies and applied to partially distinct cohorts, arrived at the same result reinforces the evidence-based conclusion that there is no time-related drift in GBM outcomes. It is also worth noting that both our study and the one by Rahman et al. are based on clinical trials that may have enrolled patients who, under the most recent World Health Organization⁴² classification, would no longer be categorized as having GBM, such as those with WHO grade 4 astrocytoma, IDH-mutant. While this reclassification did not directly affect the consistency of our conclusions, since all included studies and ones used for the validation concluded the enrollment prior to the revised classification,^{21,22,34–36} one of the included studies²² excluded patients with IDH-mutant glioma. However, a set of sensitivity analyses demonstrated that our conclusion regarding the absence of any time-trend drift in patient prognosis was not influenced by this single study. At the same time, it is important to emphasize that our conclusions regarding time-trend bias refer to the previous classification system. The updated classification will allow for a more granular evaluation of time-trend bias in the future, better aligned with the current understanding of these distinct glioma entities.

In conclusion, our results challenge the concept of time-trend bias in neuro-oncology, a limitation frequently cited in discussions about the applicability of external controls. While this can alleviate concerns regarding external controls within the framework of well-defined statistical considerations outlined in the referenced review,⁶ it also presents a disquieting outlook on GBM clinical research. Indeed, after accounting for the main GBM prognostic factors, no significant improvement in GBM patient prognosis has been observed over the past 2 decades within the homogeneous context of Phase III clinical trials.

Supplementary material

Supplementary material is available online at Neuro-Oncology (https://academic.oup.com/neuro-oncology).

Funding

No specific funding sources were used for this work.

Conflict of interest statement

To the best of our knowledge, none of the authors has conflicts of interest related to the findings or materials presented in this study.

Author Contributions

Study conception and design: G.S.; data analysis: G.S., K.D., and A.C.C.; manuscript drafting: G.S., K.D., and A.C.C.; manuscript revision and final approval: G.F.

Data Availability

The analyses reported in this study were conducted using the individual patient pseudo-data database previously published as Supplementary Material in Sferruzza et al.⁹ The code used for the analysis will be made available upon reasonable request.