Longer-term (≥ 2 years) survival in patients with glioblastoma in population-based studies pre- and post-2005: a systematic review and meta-analysis

Translation of survival benefits observed in glioblastoma clinical trials to populations and to longer-term survival remains uncertain. We aimed to assess if ≥ 2-year survival has changed in relation to the trial of radiotherapy plus concomitant and adjuvant temozolomide published in 2005. We searched MEDLINE and Embase for population-based studies with ≥ 50 patients published after 2002 reporting survival at ≥ 2 years following glioblastoma diagnosis. Primary endpoints were survival at 2-, 3- and 5-years stratified by recruitment period. We meta-analysed survival estimates using a random effects model stratified according to whether recruitment ended before 2005 (earlier) or started during or after 2005 (later). PROSPERO registration number CRD42019130035. Twenty-three populations from 63 potentially eligible studies contributed to the meta-analyses. Pooled 2-year overall survival estimates for the earlier and later study periods were 9% (95% confidence interval [CI] 6–12%; n/N = 1,488/17,507) and 18% (95% CI 14–22%; n/N = 5,670/32,390), respectively. Similarly, pooled 3-year survival estimates increased from 4% (95% CI 2–6%; n/N = 325/10,556) to 11% (95% CI 9–14%; n/N = 1900/16,397). One study with a within-population comparison showed similar improvement in survival among the older population. Pooled 5-year survival estimates were 3% (95% CI 1–5%; n/N = 401/14,919) and 4% (95% CI 2–5%; n/N = 1,291/28,748) for the earlier and later periods, respectively. Meta-analyses of real-world data suggested a doubling of 2- and 3-year survival in glioblastoma patients since 2005. However, 5-year survival remains poor with no apparent improvement. Detailed clinically annotated population-based data and further molecular characterization of longer-term survivors may explain the unchanged survival beyond 5 years.

Information sources and search. We searched MEDLINE and Embase on 15 March 2019 using a combination of search terms for glioma or glioblastoma, population-based, registry, and survival or mortality (Supplementary material Appendix 1). We screened bibliographies of eligible studies for any studies missed by the electronic searches.
Study selection. After removing duplicate records, two reviewers (MTCP and PMB) screened all titles and abstracts. Studies passing the initial screening underwent full eligibility assessment against the inclusion and exclusion criteria. We resolved any disagreements or uncertainties through discussion between the authors. Data collection process. We developed a data collection tool using a subset of eligible studies. One reviewer (MTCP) extracted data from eligible studies against the data collection tool. A second reviewer (PMB or JDF) resolved uncertainties by discussion.
Data items. We collected data on study characteristics including year of publication, country of population, registry or database used, recruitment period, any study-specific selection criteria, total sample size, median age, gender, number and proportion of patients receiving different treatments (surgical resection, chemotherapy, radiotherapy), and number and proportion of patients surviving at 2, 3, 5, and 10 years. As cancer registries generally do not routinely collect the specific chemotherapy drug and its regimen, we could not quantify the proportion of patients receiving multimodal therapy. If a study reported survival data stratified by recruitment period, we extracted the stratified survival data. Therefore, a study may contribute to meta-analyses of more than one recruitment period. When studies only reported survival data on a graph, we used Plot Digitizer (https :// plotd igiti zer.sourc eforg e.net/) to extract data.

Risk of bias assessment.
As there is no published risk of bias tool specifically for evaluating populationbased observational studies without a pre-defined exposure, we assessed risk of bias based on the preliminary ROBINS-E tool developed by GRADE 7,8 . The assessment aimed to evaluate the risk of bias affecting the survival estimate in the study. Categories assessed for bias were patient selection, diagnostic certainty, handling of missing data, and outcome measurement (Supplementary Table S1). Based on risk of bias assessment in each category, we rated the overall risk of bias as low, moderate, serious, or critical. Summary measures. Summary measures were proportions of patients with glioblastoma surviving at 2, 3, 5, and 10 years.

Synthesis of results.
We stratified studies according to whether the recruitment period was before 2005 or during or after 2005 as a proxy to wider use of multimodal treatment in the later period, informed by the 2005 landmark trial. This dichotomy does not imply a definite use of multimodal treatment since not all patients received this treatment and clinical practice of offering this treatment varies. If several publications described the same population, we included the largest sample size. Our target population was all individuals diagnosed with glioblastoma, therefore we excluded studies with selected cohorts-for example, additional exclusion criteria based on age or treatment limit-from the meta-analyses. We meta-analysed survival data at 2, 3, and 5 years from population-based studies stratified by recruitment period using Stata 16.0 (StataCorp) using a randomeffects model. We quantified heterogeneity using the I 2 statistic.
Additional analyses. By analysing non-overlapping patients stratified by recruitment period, we inevitably excluded some studies from the meta-analyses. Additional analyses of these excluded studies with recruitment period starting before and ending after 2005 allowed us to compare their pooled estimates with the primary analyses. Sensitivity analyses restricting to low risk of bias studies assessed how this affected the meta-analyses.
We observed that survival estimates from East Asian studies appeared higher than European and North American studies. To assess the contribution of the East Asian studies to the heterogeneity observed, we performed sensitivity analyses by meta-analysing the same survival data without studies of East Asian populations. www.nature.com/scientificreports/ Further sensitivity analyses included moving the proxy year to 2006 to represent the time-point after which multimodal treatment was likely to be more common. We also meta-analysed studies where patient recruitment started before and ended after 2005. These studies contained some patients also included within populations contributed to the main analyses. However, no patients were included more than once in this additional meta-analysis.

Results
Study selection and characteristics. Our search strategy retrieved 503 records. There were 135 potentially eligible studies, of which 63 were included in this review after full eligibility assessment ( Fig. 1) 9-37 . One study reported two distinct populations 14 . Therefore, there were 63 eligible studies representing 64 populations from 17 countries that reported overall survival at ≥ 2 year in patients with glioblastoma. Detailed study characteristics are presented in Supplementary Table S2 (p 4-7). The United States was the most frequently reported population (n = 29; 45%). Twenty-seven studies reported survival estimates in Western European populations (including Switzerland and the United Kingdom), and three in Canada. Of the remaining studies, four were from East Asian populations and one from Australia. The median length of recruitment period was 8.5 years (IQR 5-14 years). Median age ranged from 58 to 75 years. In studies that did not use the predominantly male Veterans Health Administration (VA) database, the percentage of females ranged from 32 to 61%. The percentage of patients receiving each of the following treatments ranged from 25 to 100% for resective surgery; 39-100% for radiotherapy; 7-100% for chemotherapy. There were 55 (87%) studies reporting 2-year survival, 31 (49%) reporting 3-year survival, 32 (51%) reporting 5-year survival, and 3 (5%) reporting 10-year survival.

Risk of bias.
Based on our risk of bias assessment, 29 (46%) studies had low risk of bias (Supplementary   Table S2). Nine studies included glioblastoma cases without histological confirmation. Fourteen studies (22%) had a sample size of less than 300. Demographic information (age and gender) were provided in 30 (48%) studies. Only 21 (33%) studies reported the proportion of patients receiving each of resective surgery, radiotherapy, and chemotherapy.  10-year survival. Only 3 (5%) studies reported 10-year survival 29,33,37 . Two studies used the United States SEER database 33,37 , while one was from Canada using a national cancer registry 37 . The recruitment periods

Sensitivity analyses.
Restricting quantitative analyses to low risk-of-bias studies showed the respective pooled survival estimates for the earlier and later periods were: 11% and 18% at 2 years; 5% and 8% at 3 years; and 3% and 3% at 5 years ( Supplementary Fig. S1). Because survival estimates from East Asian populations were consistently higher than the pooled estimate, we carried out sensitivity analyses excluding the East Asian studies from meta-analyses. The respective pooled survival estimates without the East Asian populations for the earlier and later periods were: 8% and 16% at 2 years; 2% and 11% at 3 years; and 2% and 4% at 5 years ( Supplementary  Fig. S2). Instead of using year 2005 as a proxy time-point after which wider use of multimodal therapy occurred, we stratified studies by their recruitment period into before 2006, and during or after 2006. There was no change to the number of studies available for meta-analyses in the earlier period. In the later period, meta-analyses at 2, 3 and 5 years had 5, 3, and 1 fewer studies, respectively. The respective pooled survival estimates for the earlier and later period using 2006 as the proxy time-point were: 9% and 19% at 2 years; 4% and 12% at 3 years, and 3% and 4% at 5 years ( Supplementary Fig. S3).
Meta-analyses of studies that did not contribute to the main analyses are presented in Supplementary Material (Fig. S4). These studies contain some overlapping patients with populations in the main analyses. The pooled  Proportion surviving 2 years www.nature.com/scientificreports/ estimates of the proportion of patients surviving at 2 and 3 years were 12% and 8%, respectively. Both were between the pooled estimates for the earlier and later periods. At 5 years, the pooled estimate was 4%.
Older populations. Regardless of recruitment period, there were five studies that restricted their cohorts to patients ≥ 65 years 20,27,34 and ≥ 70 years of age 23,36 . Populations in these studies were based in the United Statesfour from SEER 20,27,34,36 and one from the National Cancer Database (NCDB) 23 . The 2-year survival estimates ranged from 2 20 to 11% 23 , and 3-year survival ranged from 0.4 20,36 to 5% 23 . The 5-year survival was 0.2% in the only study reporting at this time point 20 . One study had within-population comparison for the earlier and later recruitment periods and reported an increase in survival estimate at 2 and 3 years from 2 to 7% and from 0.9 to 3%, respectively 36 .

Discussion
This systematic review of 63 population-based studies of which 22 were included in the primary analyses found that 2-year survival was 18%, 3-year survival was 11%, and 5-year survival was 4% in patients diagnosed with glioblastoma during or after 2005. There was a doubling of 2-and 3-year survival in contemporary populations compared with patients treated prior to 2005. These survival gains were also evident in older (≥ 65 years) patients. However, there was no apparent improvement in 5-year survival. The most significant development in clinical management of glioblastoma in the past two decades has been the landmark trial of radiotherapy plus concomitant and adjuvant temozolomide 3 . This clinical trial among 573 patients with a median follow-up of 28 months demonstrated that when compared with radiotherapy alone, the combination therapy improved the median survival from 12.1 to 14.6 months. The reported 2-year survival was 8% in the radiotherapy alone group, and 20% in the combination therapy group. These survival probabilities are very similar to our findings for 2-year survival from the meta-analyses of earlier and later time periods. Withinpopulation comparisons, which should be less prone to confounding, confirmed a higher survival in the later versus earlier period. Although it is not possible to directly examine whether this improvement is secondary to the use of multimodal therapy, there was a rapid increase in the use of adjuvant and concomitant temozolomide  Proportion surviving 3 years www.nature.com/scientificreports/ with radiotherapy after the trial publication 38 , at least in the United States. The European 39,40 and the United States 41 neuro-oncology guidelines have established multimodal treatment for glioblastoma as the standard of care in many countries. Hence, it is possible that the increased use of multimodal therapy contributed to the improvements observed. Our findings also demonstrate the value of using health data to assess the impact of therapeutic advances on a disease at a population level. This approach can evaluate the real-world risk-benefit ratio of treatment at scale for treatments that have significant side-effects.
International variations in incidence of malignant astrocytoma 42 and overall survival, may suggest different aetiologies, underlying disease processes, or risk exposure such as ethnicity 1 . Investigations into these factors may elucidate different underlying pathological mechanisms. Differences in cancer registration practice could also contribute to survival variation. The International Cancer Benchmarking Partnership (ICBP), which comprises 13 jurisdictions from 6 countries, has reported differences in capturing incidence date 43 . Furthermore, the increased use of brain imaging 44 and high prevalence of incidental findings 45 may further contribute to the lead-time bias.
There has been significant progress in understanding genomic drivers implicated in glioma tumorigenesis. Multidimensional data from genomics studies such as The Cancer Genome Atlas (TCGA) 46 , REMBRANDT study 47 , and the Chinese Glioma Genome Atlas (CGGA) 48 have identified molecular signatures associated with glioma subtypes and their prognoses 49 . Isocitrate dehydrogenase (IDH) mutation and 1p/19q co-deletion are most prominent markers associated with a more favourable prognosis 49 . The importance of these markers are reflected in the 2016 update of the WHO classification of central nervous system (CNS) tumours 50 , which incorporates molecular markers into the definition of tumour entities, in addition to histological features. Future studies should examine outcomes and therapy responses amongst these molecular subtypes.
Studying clinical phenotypes can gain knowledge of glioblastoma disease mechanisms. In a recent study using radiographical and transcriptomics data, it was suggested that female and male patients with glioblastoma have different molecular mechanisms 51   Proportion surviving 5 years www.nature.com/scientificreports/ studies of long-term survivors are challenging given that there are very few subjects, with a varying degree of clinical detail. The observation that the proportion of patients surviving 5 years has not changed, may indicate a molecularly distinct group of patients within this heterogeneous disease. In order to clarify this hypothesis, future studies should include comprehensive clinical details with molecular markers in a large patient cohort. This systematic review and meta-analysis presented international real-world data drawn from populationbased studies. We evaluated longer-term survival beyond the median survival observed in the landmark trial, and we were able to compare survival in the era of contemporary management to an earlier period. Our data also addressed the survival uncertainty in the older population. However, there were some limitations to our study. A high proportion of eligible studies used the same national database. We adopted a strategy to prevent any patient from being included more than once in our analyses. Inevitably this led to the omission of patients recruited in some time periods. However, our sensitivity analyses comparing the pooled estimate of these omitted studies to that of the main analyses showed a confirmatory trend of survival improvement (Supplementary Fig. S4). There was considerable between-study heterogeneity despite excluding studies with a high risk of bias and examining heterogeneity in our sensitivity analyses. Unmeasured factors such as unrecorded variables, missing treatment data, and variations in data coding would contribute to the variation of survival estimates 57 . Patients in some studies did not have histological confirmation; patients with better prognosis associated with lower grade gliomas or other brain tumours may have been included. This ascertainment bias is unlikely to have significantly affected our findings since studies in the meta-analyses, where some patients did not have histological confirmation, had similar survival to other studies. While this review examines the trend in survival in relation to a landmark clinical trial, we were unable to evaluate directly the pattern of treatment and its impact on survival due to limited treatment information available from eligible studies. Within-population comparison of different time periods, which may partially control for unmeasured factors, was only possible in a few eligible studies. We are, therefore, unable to directly attribute the survival trend observed to a change in treatment patterns. We used the year of 2005 as a proxy time-point for the introduction of wider use of multimodal therapy. Although some studies specified the more prevalent use of multimodal therapy, not all would have taken up recommendations from the clinical trial. However, our sensitivity analysis showed no substantial change in trend and having the cut-off year at 2005 allowed more studies to be included ( Supplementary Fig. S3). Survival improvement may be attributable to other factors in addition to the multimodal treatment, such as more prevalent use of neuroimaging leading to earlier diagnosis, strengthening of multidisciplinary neuro-oncology care, and advances in neurosurgical techniques.

Conclusions
Overall estimates of survival among patients with glioblastoma have at least doubled since 2005 to 18% at 2 years and 11% at 3 years. This may reflect treatment response to modern therapeutic approaches. However, longer term survival remains poor and there appears to be a lack of improvement in 5-year survival. Large populationbased studies with detailed clinical characteristics would clarify whether there has been an increase in the use of multimodal therapy and whether multimodal therapy affects survival beyond 5 years. Differences in msolecular markers of tumour subtypes may shed light on the underlying disease mechanisms that influence survival.