Identifying appropriate external control datasets in support of future glioblastoma clinical trials leveraging external data

Abstract

Background

Recent interest in leveraging external data for clinical trial design and analysis in glioblastoma has raised questions on the identification of appropriate data to use as external controls for future trials. We perform a comprehensive analysis assessing candidate sources of external data and comparing clinical trial and real-world datasets in newly diagnosed glioblastoma.

Methods

Individual patient-level data (PLD) from several clinical trials, a large academic institutional database and a registry (National Cancer Database) were used for analysis of patients receiving standard of care radiation with concurrent and adjuvant temozolomide. Data summaries from randomized trials 2012–2022 were analyzed to account for trials without available PLD. Multivariable modeling was employed to compare survival across datasets.

Results

In total, 8 datasets with PLD for 3061 patients with newly diagnosed glioblastoma treated with standard chemoradiation were analyzed. Patients on trials were younger (age < 60:64% vs. 48%, p < 0.001) and had higher KPS (KPS≥90:58% vs. 48%, P < .001) compared to non-trial patients. Patients in clinical trials exhibited inferior survival relative to non-trial patients (HR 1.30,95%CI 1.13-1.48, P < .001) after adjustment for age, sex, KPS, extent of resection and MGMT methylation status. In assessment of data summaries of 19 randomized trials, there was no detectable time-trend toward improved outcomes 2012-2022.

Conclusions

In newly diagnosed glioblastoma patients treated with standard of care chemoradiation, there were significant differences between trial datasets and real-world datasets but no evidence of a trial effect benefit from trial participation. After adjustment of relevant covariates, there was no evidence of temporal drift of improved survival over the last decade.

Glioblastoma (GBM), the most common primary adult brain tumor in adults, has a dismal prognosis despite standard of care therapy of surgery, radiation therapy and chemotherapy.¹ Drug development is slow, expensive and has yielded limited advances in recent decades.^2,3 In the context of this unmet need, the use of external data in neuro-oncology clinical trials has recently gained interest, with the potential for increased efficiency of future studies by leveraging external data for decision making during trials (eg, early stopping⁴) and final analyses.^5–7 In the last five years a number of innovative methodologies to leverage external data have been developed in the statistical literature. The FDA has acknowledged and provided guidelines for such novel designs in 2023.⁸ External data can come from previously completed clinical trials and real-world data (RWD), defined by the FDA as data related to patient health status or delivery of health care routinely collected from a variety of sources.⁹ The use of externally augmented clinical trial (EACT) designs, which incorporate data external to the trial, is appealing in GBM given its relative rarity, ineffective standard of care therapies, frequent use of single arm trial designs,¹⁰ emerging classes of treatments developed through small studies,¹¹ patient reluctance to randomization and availability of several datasets of patients treated with the current standard of care.¹²

The National Brain Tumor Society hosted a Research Roundtable on July 20, 2023 with experts in neuro-oncology, data science, biostatistics, as well as representatives from patient advocacy and the FDA. Stakeholders convened in an interdisciplinary discussion on challenges, opportunities, risks, and regulatory considerations on the integration of external data in neuro-oncology clinical trials. Several questions were identified as important to accurately assess the utility of integrating external data in future GBM clinical trials. Here we conduct a study to address two critical questions that were raised about the choice of external control datasets.

Question 1:  Integration of external datasets in future trials: should we use data from completed trials, or can we incorporate real-world data (RWD), or a combination of clinical trial datasets and RWD? A significant component of the Research Roundtable discussion focused on the identification of the most appropriate data sources that could constitute useful and adequate external controls. Clinical trial datasets are appealing as they comprise patients who meet clinical trial eligibility with high quality data annotation, and they are often preferable as an external data source.¹³ Nonetheless, RWD can be more available and there has been interest on whether RWD could be used or in combination with clinical trial data. As a part of this discussion, concerns were raised about potential trial effects, defined as a tendency of patients enrolled in clinical trials to have superior outcomes compared to non-trial patients,¹⁴ and such a trial effect could bias the analyses that incorporate RWD.¹³

Question 2:  Selection of recent external data: while contemporaneous datasets are ideal, how recent should the external data be? Which datasets should we consider obsolete in GBM? Given the lack of therapeutic advances in GBM but potential impact of other advances in the management of patients (e.g. supportive care, surgical and radiation techniques), there were questions on the appropriate temporality of data sources.¹⁵ In general, more contemporaneous datasets are ideal, but the presence and extent of a temporal trend of improved outcomes over recent years, has not been characterized in GBM.

We were interested in evaluating discrepancies across datasets (completed trials vs. RWD) and temporal trends that could introduce bias and compromise the scientific rigor of future studies by integrating external data. The data collection encompasses clinical trials, large academic institutional datasets and a large cancer registry. For clinical trials where patient-level data (PLD) was not available, we used available data summaries from publications during the last decade to provide a comprehensive multi-study analysis. In total, we perform an analysis based on several clinical trial datasets and RWD with PLD of over 3000 newly diagnosed patients.

Methods

Patient-Level Datasets

We requested PLD from newly diagnosed glioblastoma trials conducted since 2005 and obtained PLD from 6 randomized clinical trials (RCT). We performed an analysis, with PLD from 8 datasets (previous trials or RWD) including 3061 newly diagnosed patients that were provided upon investigator request (Supplementary Table 1A). We used data from patients receiving standard of care radiation therapy with concurrent and adjuvant temozolomide (Table 1). Pre-treatment patient information included age, sex, Karnofsky performance status (KPS), extent of resection, and MGMT gene promoter methylation status. With respect to outcomes, PLD included overall survival (OS) times. OS was defined as the time of death or censoring from randomization (for RCTs) or from the start of RT (for the RWD). Given that clinical trials often include the MGMT promoter methylation status as eligibility criteria, we conducted dedicated analyses of PLD for MGMT methylated and unmethylated subpopulations. This study involved re-analysis of datasets that have been previously published upon, and the study was conducted under approved protocol by the Dana-Farber/Harvard Cancer Center institutional review board.

Matching Analysis

Propensity-score based methods were used to compare the survival distributions of patients treated with chemoradiation across trial and RWD datasets. We selected the distribution of pre-treatment patient characteristics in the RWD-DFCI as reference distribution. We then used propensity score-based matching to compare and graph survival for the remaining studies (one at the time). These analyses incorporate adjustments to account for differences between pre-treatment patient profiles in the DFCI and in the other studies.¹⁶

For each GBM dataset k = 1,..K other than the RWD-DFCI dataset, we merged the dataset k and the RWE-DFCI into a single data matrix and added an additional indicator variable A that indicates whether a specific patient in the dataset belongs to the RWD-DFCI (A = 1) or not (A = 0). We then fitted a binary regression model to the merged data, with A as the dependent binary variable and pre-treatment variables (ie age, sex, KPS > 80, extend of resection, MGMT methylation status) as predictors. All except one (PMID22120301) dataset categorized extent of resection as gross total resection (GTR), subtotal resection (STR), and biopsy. We estimate the effect of extent of resection on OS with STR designated as reference. We use the logit-model to estimate for each patient i in the merged dataset a propensity score $ei$ (i.e. an estimate of p(A = 1|X)). We then match each patient i = 1, . . . , n in the RWD-DFCI to a patient $ri$ in the GBM dataset k with similar propensity score (PS). Specifically, for each patient i RWD-DFCI, we determined the set of all patients j in dataset k whose PSs differ from by less than 0.2 multiplied by the standard deviation of propensity scores. If this set is empty no match was obtained, otherwise we select the patient $ri$ in dataset k with the smallest difference $|ei−eri|$⁠. We then computed a Kaplan-Meier survival curve using the outcome data of the matched patients.

Confidence intervals for the matched survival probabilities and the medians were obtained via bootstrap resampling of the pre-matched data. Specifically, for a fixed dataset (for instance, the NCDB data), we resampled c = 1, . . . , 1000 times, with replacement, the RWD-NCDB and RWD-DFCI data. For each of the c = 1, . . . , 1000 pairs of RWD-NCDB and DFCI datasets, we repeated the matching analysis, obtained matched KM curves and the corresponding median $mc$ for the RWD-NCDB data. We then used the bootstrap sample ${mc}c=11000$ to obtain a 95% confidence interval for the matched median OS in the RWD-NCDB. The width of the interval is equal to the difference between the 2.5% and 97.5% sample quantiles of ${mc}c=11000$⁠.

Multi-Variable Cox Proportional Hazards Analyses

In addition to the Kaplan–Meier analysis after matching described in the previous subsection, we estimated multi-variable Cox proportional hazards models to quantify and test (i) potential cross-study variations of the conditional outcome distributions (given pre-treatment variables) across GBM datasets and (ii) potential differences in the outcome distribution of RWD and RCT datasets. Specifically, we fitted Cox models to the combined data with conditional hazard rates for patients in each dataset k,

with a vector of pre-treatment patient characteristics $x$⁠, covariate effects b, and $I(kisRWD)$ is one if study k is a RWD and zero otherwise, study effect $uk$⁠. The study effects $(uk)$ for all RWD and RCT datasets are assumed to follow a normal distribution with mean 0 and variance $σu2$⁠, which we estimate from the data. We used a Wald test to evaluate whether the conditional outcome distributions, given pre-treatment characteristics, differ between RWD and RCT studies, $H0:νRWE=0$⁠.

To assess the hypothesis of homogeneity (ie, identical conditional outcome distributions across trials and across RWD datasets) we test $H0:σu2=0$⁠. In particular, we fitted a simplified Cox model, setting $u1=…=uk…=0$⁠. We then used a likelihood ratio test to evaluate if the random-effects model eq. (1) fitted the data significantly better than the fixed effects model $h(t|x)=h0(t)exp{νRWEI(kisRWD)+x′b}$ without the study-specific random effects.

We conducted additional landmark analyses to evaluate the sensitivity of tests and model estimates to the definition of the outcome variables. Specifically, we re-estimated the random-effects and fixed-effects Cox models in a subset of patients that survived at least z months after the initiation of radiation treatment. We used z = 2, 3, 4 months.

We also assessed study effects of each clinical trial relative to other trials among MGMT-agnostic, MGMT methylated and MGMT unmethylated datasets. For each study k we fitted a fixed effects Cox PH model $hk(t|x,zk)=h0(t)exp{νkzk+x′b}$ where for each patient in the MGMT-agnostic (MGMT methylated or MGMT unmethylated) datasets $zk=1$ if the patient belongs to study k and $zk=0$ if the patient belongs to any other study. For example, in the MGMT methylated subset, NCT00689221 is compared using a Cox PH model to NCT00943826, NCT00441142 and PubMed22120301. We then used a Wald-test to evaluate $H0:vk=0$⁠.

Summary Level Data from Glioblastoma Clinical Trials Over the Last 10 Years

In order to provide a comprehensive analysis of randomized GBM trials, we searched published randomized GBM trials. We performed a PubMed search on November 11, 2022 with the keyword of “glioblastoma” for clinical trials published over the last 10 years for newly diagnosed GBM. Included publications reported on a randomized clinical trial evaluating a therapeutic intervention with a time-to-event outcome and Kaplan–Meier survival curve with an at-risk table (Supplementary Table 1B). All included publications were reviewed to extract relevant summary patient population characteristics.

Summary Level Data Analysis (to Increase the Data Collection Beyond Studies With Accessible PLD)

We extracted available pre-treatment patient summary information, trial characteristics (sample size, data of first and last enrollment, median age, sex, % KPS >=90, etc.,) and outcome summaries (median OS and reported 95% CI) for the TMZ + RT arms reported in GBM RCT publications identified via the PubMed search (Table 2). For each of these GBM RCT publications we used data-imputation to impute arm-specific covariate summaries not reported in the publication of the trial (Supplementary Table 2).

To investigate temporal trends, for each k = 1, . . . , 19 RCT publications, we used the remaining 18 out of 19 publications to estimate a regression model with reported log-median RT/TMZ OS as dependent variable, and aggregated pre-treatment patient summaries as predictors. We then use this model to predict the median OS of the control arm in study k based on the study-specific aggregated patient summaries. These 19 repeated analyses are used to graph point for each study k: (i) on the x-axis the time (averaging) of enrollment of each trial and (ii) on the y axis the residual (observed—predicted) study-specific median OS based on a model trained by the other 18 datasets. Note that in presence of a temporal trend towards improved median OS in recent studies (after adjusting for the composition of the enrolled patients) we expect the same trend to persist graphing the residuals.

The 19 prediction models in these analyses included summaries of the control arms (a 6-dimensional vector per study: size of the control arm, median age, proportion of males, extent of resection (%), KPS>=90 (%), Unmethylated MGMT status (%)) and OS median (in log scale) of the studies as dependent variable. Control arm size was included as it can correlate with variance of the Kaplan-Meier estimate of the median and possibly other unmeasured confounders. The models did not include information on the enrollment period of each study. To summarize, we investigate if the residuals (observed median OS vs. fitted value) exhibit a trend, with larger residuals for the more recent studies or vice versa.

Results

Analysis of Individual Patient-Level Data

Pre-treatment individual profiles

.—A summary of patient characteristics of RCT (NCT00441142, NCT00689221, NCT00813943, PMID22120301, NCT00943826, NCT02977780) and RWD datasets (RWD-DFCI, RWD-NCDB) are provided (Table 1). All patients received standard of care radiation therapy with concurrent and adjuvant TMZ. Of note, patients on clinical trials were younger (median age 57 vs. 61, P < .001) and had better KPS (KPS≥90 58% vs. 48%, P < .001) relative to patients from RWD sources (Figure 1, Supplementary Table 3). Differences in age and KPS between trial and RWD sources persisted for the subgroup of patients with a favorable clinical profile (age < 70, KPS > 60 with a gross total resection; Supplementary Table 4). There were more MGMT unmethylated patients from RWD datasets (55%) relative to clinical trial datasets (45%) (P = .0079). Further descriptive summaries on relevant clinical covariates across datasets are available in the supplement (Supplementary Figure 1).

Survival distributions, estimated using propensity scores, are plotted in Figure 2, including subsets of MGMT methylated patients only (Figure 2B), and MGMT unmethylated patients only (Figure 2C). The Kaplan-Meier estimates suggest better conditional survival distributions in the RWD datasets compared to the RCT datasets. Supplementary Figure 2 indicates that the results of the matching analysis remain nearly identical when we use the distribution of pre-treatment variables in the RWD-NCDB dataset as reference for the matching algorithm instead of the RWD-DFCI.

The conditional OS tend to be superior in RWD compared to clinical trials.—We performed a Cox random effect multivariable model. We used a likelihood ratio test to assess if the random effects in (model 1) were needed (ie, H0: $σu2=0$ for the random effects in (1)), and the resulting P-value was not-significant (P = 0.74), supporting the use of a fixed effects model. The final fixed-effects Cox model (1) without study-effects (⁠$uk$ = 0) assumes that the effects of pre-treatment characteristics (ie, age, sex, EOR, etc.) on OS do not vary across datasets k. Therefore, larger datasets, have larger contributions to the likelihood-based estimation of effects of pre-treatment characteristics. We also fitted a more flexible model $hk(t|x)=h0(t)exp{uk+νRWEI(kisRWE)+xbk},$ thus allowing for study-specific effects of relevant pre-treatment variables. A Likelihood ratio test did not support the use of this more complex model over the simpler model (ie, $bk=b$⁠) without study-covariate interactions.

Next, we performed a multivariable fixed effects Cox model for OS to evaluate differences based on the data source (clinical trial dataset vs. RWD, model 1 with random effects $uk$ equal zero). Residual analysis (using martingale and Schoenfeld residuals) for the final fixed-effect model did not indicate any violations of PH model assumption. In this analysis (Table 3), patients on clinical trial datasets had inferior survival compared to patients in RWD datasets (adjusted hazard ratio [AHR] 1.30, 95%CI 1.13–1.48, P < .001) after adjustment for sex (male AHR 1.12, 95%CI 1.00–1.25, P = .036), KPS≥90(AHR 0.67,95%CI 0.59–0.73, P < .001), older age(continuous variable, AHR1.03, 95%CI 1.02–1.03, P < .001), extent of resection (GTR AHR 0.84, 95%CI 0.70–1.01, P = .06), and MGMT promoter methylation (AHR0.48, 95%CI 0.43-0.54, P < .001). The Cox PH analysis was repeated using the RWD (RWD-DFCI and RWD-NCDB) and phase III data (NCT00943826 and NCT00689221). We observed the similar differences between RWD and RCT data (Supplementary Figure 3).

When we focused on MGMT unmethylated tumors, we identified a similar difference in OS based on PLD from clinical trial datasets compared to RWD (AHR 1.32, 95% CI 1.16–1.52, P < .001). Age, sex, KPS, and extent of resection remained significantly associated with overall survival in this subset of patient cohorts. Among patients with MGMT methylated tumors (Table 3), the difference in OS between PLD from clinical trial datasets compared to RWD sources was not significant (AHR 1.12, 95% CI 0.94–1.33, P = .19).

Given potential concerns about bias due to slight differences in the definitions of OS across datasets (ie, RWD: survival defined from time of start of radiation therapy; RCT: survival defined from time of trial registration), we did a separate landmark analysis including patients with OS > 120 days with a fixed effect Cox model. This produced results that were concordant with the original analysis (Supplementary Table 5). Note that the bias effect of the outlined differences in OS definitions is towards an underestimation of survival in RWD relative to trials, but in contrast we estimate a difference of opposite sign.

Similar results were obtained with landmark analysis with cut-offs at 30 days and 60 days. The landmark analysis suggested that results were stable to slight differences in the definition of the endpoint. The survival analyses identify shorter OS in trial control arms compared to RWD datasets. Based upon above, we do not associate this difference to the definitions of the indexing time, from which one can expect a difference between RCTs and RWD of opposite directionality. Similar results were also obtained in analysis focusing upon phase 3 RCT data compared to RWD (Supplementary Table 6).

We also assessed effects of each clinical trial relative to other trials among MGMT-agnostic, MGMT methylated and MGMT unmethylated datasets. In these three subsets, there are no significant effects of one trial relative to the others (each 95% CI of the hazard ratio overlap with ~1; Supplementary Table 7). There was no evidence of study-to-study variations of OS conditional on prognostic variables across clinical trials.

Analysis of Summary Level Data

We used summary level data from randomized GBM trial publications over a ten-year period. Information about these clinical trials is summarized in Table 2, and imputation was used for missing data or predictors published in non-standard formats in the publications. Prediction models (linear regression) were fitted for each leave-one-trial-out replicate to predict median survival based upon published summaries (average age, sex, MGMT methylation status, extent of resection, performance status, and size of the control arm) and summary level data from the other publications (see Methods). Predictions and 95% CI were computed for each trial.

The leave-one-trial-out predictions of survival are graphed alongside the actual median survival times reported in each publication, and with confidence intervals (publications were ordered on the x-axes according to average calendar date of enrolment), in Figure 3A. We do not observe any noticeable time trend for the reported median OS. We evaluated potential temporal drifts of outcomes in GBM, plotting the residuals (difference of observed median survival time—predicted value) against the central trial-specific year of enrollment. Prediction models did not include information on the enrollment periods. There was no evidence of time trend in survival over the studies in a ten-year period (Figure 3B).

Discussion

The interest in the use of external datasets to design trials and develop new treatments in neuro-oncology have raised important questions about the comparability of different data sources. Different sources of data can provide distinct advantages and risks when considering their use for EACT designs.^13,17 We provide a comprehensive analysis of patient-level data sources in newly diagnosed GBM patients who were treated with standard of care therapy, with radiation and concurrent adjuvant temozolomide. In addition to broad inclusivity of datasets, data was generally well annotated with respect to pre-treatment clinical covariates that are essential for survival analyses in GBM.¹² With respect to pre-treatment co-variates, unsurprisingly, patients in clinical trial datasets were younger and had higher KPS (Figure 1). While there has been interest in broadening clinical trial eligibility,¹⁸ clinical trials tend to enroll patients healthier than the general population in part through stringent eligibility criteria which may limit generalizability of trial results. With adjustment of clinical covariates, however, we detected inferior outcomes of clinical trial dataset patients compared to the RWD, an unanticipated result. These findings suggest a need for caution if trying to interchangeably use clinical trial and RWD datasets when implementing EACT designs.

The reasons why we observe inferior survival across clinical trial datasets compared to RWD sources is unclear, but several possibilities could be contributing. While we did not observe large differences between the censoring rates and the median follow-up times of the RWD and trial data, our analysis involves censoring and the so called non-informative censoring assumption (conditional independence of censoring and outcomes given pre-treatment variables in each study). The observed difference in OS between RWD and trial data might be associated with a bias mechanism due to violations of the non-informative censoring assumption, which is nearly ubiquitous in survival analyses. For example, RWD from large academic medical centers may include patients who are preferentially lost to follow-up (e.g. too sick to continue follow-up). With respect to clinical trial datasets, patients may be prone to more delays to treatment initiation due to protocols and screening, though short delays in chemoradiation initiation have not been associated with a relevant change in outcomes.¹⁹ Further study would be important to better understand if these differences explain the discrepancies between outcomes distributions (conditional on pre-treatment clinical profiles) in clinical trial datasets and RWD. Since data collection and bias mechanisms likely explain the difference in outcomes between RCT and RWD sources, our findings should not dissuade the need to incentivize clinical trial enrollment for GBM patients, as this is an important step towards better therapies and improved outcomes for patients.

Given that clinical trial datasets provide clearly defined populations and accurate clinical annotation,²⁰ these datasets may be well suited for use in EACT designs. To better understand possible risks, we evaluated possible study effects across clinical trial datasets. We did not detect significant study-to-study variations across clinical trial datasets after adjustment of clinical covariates. Our results support the use of clinical trial datasets, ideally with recent studies if available. While RWD remains appealing given its availability and potential for ease of access and contemporality, further investigation is warranted to identify effective approaches—ranging from data harmonization, procedures to scrutinize data capture mechanisms, and statistical modeling—to leverage such RWD for EACT designs. Standardization of reporting of clinical variables across neuro-oncology is crucial for improved harmonization of RWD and clinical trial datasets. Until the transition towards higher quality RWD occur, clinical trial data may be preferable given their known advantages (e.g. well-defined trial population, standard definitions of endpoints), though efforts and resources to share usable control data from recently completed randomized GBM trials are needed.

While many oncologists believe that cancer patients who enroll in clinical trials may have better outcomes compared to patients treated outside of a clinical trial,²¹ scant evidence exists on whether trial participation itself affects outcomes among patients receiving standard of care therapy in a randomized trial.²² Our results suggest that there is no such beneficial clinical trial effect in newly diagnosed GBM after accounting for possible confounding clinical covariates. Our results are in contrast to a few single institutional analyses which were smaller studies, some without adjustments for all relevant clinical covariates (eg, MGMT promoter methylation).^23,24

With the exception of a modest benefit associated with tumor-treating-fields,²⁵ there have been minimal improvements in therapies for GBM over the last two decades since the EORTC/NCI trial establishing the role of concurrent and adjuvant temozolomide.²⁶ Despite a lack of therapeutic advances, there has been speculation that there may be improvement of outcomes over time, possibly due to better supportive care, surgical or radiation techniques.²⁷ To address this, we used summary level data across published clinical trials over the last 10 years to inspect the possibility of temporal drift effects. We did not detect improvement of clinical trial outcomes of clinical trial GBM patients after adjustment for relevant clinical covariates (Figure 3B). In part to account for modern classification of GBM (ie, confirmed IDH wild-type status),⁸ external control datasets should ideally be contemporaneous and include features such as relevant patient biomarker an classification data. Our analysis of published clinical trials did not identify trends in the conditional outcome distribution of the control regimen of radiation therapy with concurrent and adjuvant temozolomide therapy over the past ten years.

There are several limitations in this study. We did not account for tumor treating fields, which is an FDA approved therapy for GBM; tumor treating fields are excluded from most GBM trials and used in the care of a minority of the patient population. The changing classification of adult gliomas merits particular attention; IDH mutation status was not uniformly reported in older datasets (Table 1), but the definition of GBM per the 2021 WHO Classification requires IDH wild-type status.²⁸ Therefore, ideal external data sources should include IDH mutation status. Furthermore, protocols did not explicitly define variables such as extent of resection, though we presumed that gross total resection indicated removal of enhancing tumor without measurable disease on post-operative imaging across datasets. Despite requests for PLD, only a subset of requested datasets was able to be obtained. Trial datasets with PLD did not include the most recent completed glioblastoma trials, and our analysis of temporal drift effects therefore focused on the use of data summaries. The study was not able to compare outcome distributions and potential differences across geographic location or practice setting. While RWD can come from many different sources, one major RWD source in this study was a large dataset from a single large academic medical center, and the other source of RWD was the NCDB registry where 70% of patients were collected from academic centers (Supplementary Table 8). This is relevant as there can be differences in outcomes in oncology patients treated at large academic medical centers compared to smaller center.^29,30 Our summary-level analysis was based on a small number of datapoints (19 datasets) with a moderate number of predictors (P = 6). Moreover, the reporting of summary-level data varied widely across publications, which required data-imputation. These are two limitations of our analyses to estimate potential trends of trial-specific median OS of the control groups over the years.

Conclusion

Our results suggest that there should be caution in trying to use real-world data in combination with previously completed clinical trials as external data sources. Furthermore, the absence of temporal effects suggests that well-annotated clinical trial datasets from the prior decade may be appropriate for GBM trial designs leveraging external data.

Acknowledgements

This publication is based on research using data from data contributors Roche that has been made available through Vivli, Inc. Vivli has not contributed to or approved, and is not in any way responsible for, the contents of this publication. The authors thank the National Brain Tumor Society for arranging the 2023 Research Roundtable on the Use of External Control Data in Brain Tumor Clinical Trials.

Conflict of interest statement. Dr. Rahman reports advisory board consulting with Servier, Telix Pharmaceuticals; consulting with NH TherAguix; outside of submitted work. Dr. Wen reports other from Black Diamond, personal fees and other from Agios, personal fees and other from Astra Zeneca, other from Bristol Myers Squibb, personal fees and other from Eli Lilly, personal fees and other from Chimerix, other from Erasca, other from Global Coalition For Adaptive Research, other from Quadriga, personal fees and other from Merck, personal fees and other from Servier, personal fees from Anhert, personal fees from Black Diamond, personal fees from Celularity, personal fees from Day One Bio, personal fees from Genenta, personal fees from Glaxo Smith Kline, personal fees from Kintara, other from Kazia, other from MedicaNova, personal fees and other from Novartis, personal fees from Mundipharma, personal fees from Novocure, personal fees from Prelude Therapeutics, personal fees from Sagimet, personal fees from Sapience, personal fees from Symbio, personal fees from Tango, personal fees from Telix, personal fees from VBI Vaccines, outside the submitted work. All other authors have nothing to disclose.

Funding

R.R. is supported by Joint Center Radiation Therapy Foundation Grant and Kayes Technology Grant. S.V. was supported by the National Cancer Institutes (5P30CA077598-23), a DSI-Grant of the Minnesota Supercomputing Institute, and a Medtronic Faculty Fellowship. L.T. is supported by NIH Grant R01LM013352.

Authorship statement

Design and implementation: RR, SV, PYW, LT. Acquisition, analysis or interpretation of data: all authors. Writing of the manuscript, revision, approval of the final version: all authors.

Disclaimer

The National Cancer Database is a joint project of the Commission on Cancer of the American College of Surgeons and the American Cancer Society. The Commission on Cancer’s National Cancer Database and the hospitals participating in the database are the source of the deidentified data used herein; they have not verified and are not responsible for the statistical validity of the data analysis or the conclusions derived by the authors.

Data Availability

As described in the text, datasets were obtained from several different sources. Based upon existing data use agreements, data available and allowed for sharing can be available from the corresponding author upon reasonable request. Further information available at https://rconnect.dfci.harvard.edu/gbmdata/.

Prior Presentation

Study results have been previously presented in part at the American Society for Clinical Oncology 2022 Annual Meeting, Chicago, IL, USA.

References

Author notes