10 Year Glioblastoma Clinical Trial Meta-Analyses by Autonomous AI at Scale. Survival, HR, AE, and RoB scored in AI Reports and Charts, including Verifications

Question: Can glioblastoma ten year clinical trial survival times, hazard ratios, adverse events, and risk of bias from online publications be effectively pooled, scored, and visualized with insights using artificial intelligence?

Findings: Ten glioblastoma therapies from the past 10 years were compared using AI: including EF-14, EORTC 26101, CheckMate 143, and REGOMA clinical trials. Commonly used charts such as forest, funnel, and risk of bias assessment plots were generated based on human prompts. Where patient data was not available, AI utilized alternative plots or approximated individuals’ data. Diagnostic, therapeutic, and biomarker approaches were utilized to develop reports of pooled and scored metrics. Insights were visualized using OS, PFS, and Grade 3+ AE rates alongside AI generated residual heterogeneity, SCCS, and synergy scores.

Meaning: Glioblastoma tumors have a fast growth rate, are difficult to treat due to the blood-brain barrier; while only a limited number of effective therapies have been FDA approved, and are therefore often fatal. This study is an application of internet based clinical trial data that yielded rapid text and image based insights of pooled data. For instance, several insights were based on the 695 patient Stupp et al. TTFields EF-14 2017 trial (OS HR = 0.63), which had a low occurrence of high grade adverse events. On the other hand, Regorafenib was also evaluated, which had improved survival performance (OS HR = 0.50), but with 56% Grade 3-4 toxicity. The core AI experiments presented here took approximately 5.4 total hours to complete, with human verifications and paper completion within 34 days.

Design: Internet searches of top 10 glioblastoma 2015-2025 clinical trial areas utilized in standalone PRISMA-2020 aligned meta-analyses (MAs) were created by ChatGPT o3 Deep research (o3dr). For each meta-analysis, 10 charts were generated in Python by 3.7 Sonnet Extended (37se). All meta-analyses were combined into a 103,905 word dataset and processed into reports by Gemini 2.5 Pro (g25p). For each report, 37se generated 10 charts, as well as 11 charts for a 3 report combination. Grok 3 (grk3) provided efficient fixes to Python code if charts contained errors. Four standards served as templates across 27 prompt variants. Excerpts of meta-analysis and report text are included due to size, with full outputs accessible through supplementary files. Note: This study is intended for educational purposes only. 

Results: In this work, glioblastoma clinical trials such as KEYNOTE-028 and DC Vaccine were summarized into meta-analyses by AI, with best responses rates 81% of the time. Insights from ongoing Neoantigen Vaccine trials included combining personalized vaccines with checkpoint inhibitors to "prevent the exhaustion of vaccine-induced T-cells, potentially leading to deeper tumor control." AI visualized clinical trial metrics in charts with best/approximation/error responses of 30%/50%/20%. Approximations were made primarily due to lack of data, and an error was recorded if at least one mistake was observed. Report creation based on the 103,905 word dataset yielded best responses of 90%, while best quality charts of pooled or scored endpoints occurred 55% of the time, with 25% being approximated. This work represented effective text generations, while images required screening due to data and chart type complexities.

Importance: Comparing several clinical trial endpoints using grade systems, and additional intuition across four AI software manufacturers was possible using high contextual awareness and tools that were not accessible prior to early 2025. Stupp et al.’s tumor-treating fields is a widely accepted study for glioblastoma, and their results were represented consistently in multimedia throughout the study. Other treatments such as Neoantigen Vaccine and DNX-2401 + Pembro were visualized effectively with similar OS HR and SCCS metrics, and smaller patient cohorts aided by a network graph.

Conclusion: Glioblastoma 10 year clinical trial survival times, hazard ratios, and other endpoints from online publications were effectively pooled and visualized with charts across ten disease areas using AI. Additional conclusions were made combining multiple MAs with scores, as TTFields emerged "as a consistently beneficial therapy (positive SCCS, high IRBS in MGMT+), while checkpoint inhibitors (Nivolumab, Pembrolizumab) show limited efficacy (low/negative SCCS, low IRBS)." The study was made feasible by utilizing unique advantages from four different AI software manufacturers. o3dr offered ten best-in-class autonomous web search and meta-analysis generations. 37se chart generation performance featuring endpoints and scores was unmatched by other software platforms. g25p was the only model capable at a reasonable cost of processing the large meta-analysis dataset to create key insights and tables with sample calculations. grk3 proved to be fast and consistent in correcting chart errors when necessary.