New research presented at the 2026 Society for Neuro-Oncology Annual Meeting demonstrates how AI can uncover meaningful treatment patterns, improve data analysis, and unlock new value from Expanded Access Programs.
Expanded Access Programs (EAPs) serve a critical purpose: providing patients with serious or life-threatening conditions access to investigational therapies when traditional treatment options have been exhausted.
But these programs also generate something increasingly valuable to pharmaceutical companies: real-world data about how investigational therapies are used in patients who may not qualify for traditional clinical trials.
Until recently, much of this information has been difficult to analyze systematically. Patient histories, treatment modifications, safety observations, and outcomes are often distributed across multiple records, with important details buried in unstructured clinical notes.
New research presented at the 31st Society for Neuro-Oncology (SNO) Annual Meeting demonstrates how artificial intelligence can help change that.
The study, “AI-Powered Natural Language Interface for Real-World Data Analysis in Neuro-Oncology Expanded Access Programs,” was conducted by researchers affiliated with MedaSystems, Valent Technologies, and The University of Texas MD Anderson Cancer Center.
Using MedaSystems' AI-powered Model Context Protocol (MCP) technology, researchers analyzed real-world treatment data from an Expanded Access Program involving VAL-083, an investigational therapy studied in patients with glioblastoma and other difficult-to-treat neuro-oncologic malignancies.
The findings offer important insights into both the treatment experience of these patients and the emerging role of AI in real-world evidence generation.
A Closer Look at the Research
The study examined de-identified data from an Expanded Access Program for VAL-083 (dianhydrogalactitol), an investigational DNA-targeting therapy.
The dataset included:
81 patient inquiries across 16 institutions
39 patients who received treatment
Treatment histories, patient demographics, and biomarker information
Dosing adjustments, treatment cycles, and safety observations
Longitudinal patient records, including information captured in clinical notes
Researchers used the MedaSystems Valent MCP Agent to query the data through natural-language questions.
Rather than manually extracting information from individual patient records and compiling spreadsheets, researchers could ask questions about treatment duration, dosing changes, biomarker characteristics, and adverse events.
The system retrieved relevant information from the underlying records, generated analyses, and provided traceability to the original data.
Importantly, the research went beyond demonstrating that AI could answer questions. It evaluated whether the technology could produce consistent, reproducible results across different AI models and analytical approaches.
Five Important Findings from the Study
1. Treatment duration varied across patient populations
One of the most interesting observations involved differences in treatment exposure across diagnostic groups.
Among the 39 treated patients, 32 had glioblastoma, including one patient with gliosarcoma. The remaining seven had other neuro-oncologic diagnoses, including astrocytoma, ependymoma, diffuse intrinsic pontine glioma, and diffuse hemispheric glioma.
Patients with non-glioblastoma diagnoses generally remained on treatment longer, with a median of approximately 4.5–5 treatment cycles compared with two cycles for patients with glioblastoma.
Both patients with ependymoma completed five to six cycles at full dose, with no toxicity recorded.
These observations may help researchers identify patient populations worthy of further investigation.
However, treatment duration is not equivalent to clinical efficacy. Differences in disease biology, previous treatments, and concurrent therapies may have influenced the results.
The significance: AI can help researchers identify patterns across small, heterogeneous patient populations that might otherwise require extensive retrospective analysis.
2. Biomarker analysis revealed a potential research hypothesis
The study also identified an interesting pattern involving MGMT promoter methylation status, a biomarker relevant to glioblastoma treatment.
All six patients who received six or more cycles of VAL-083 had MGMT-unmethylated tumors.
In contrast, the longest recorded treatment course among patients with MGMT-methylated tumors was four cycles.
This observation does not establish a treatment benefit for patients with MGMT-unmethylated tumors. The study involved a small, nonrandomized population without a comparator group, and some patients received concurrent therapies.
Nevertheless, it illustrates how AI-assisted analysis can surface biomarker-associated patterns that warrant further evaluation.
The significance: Expanded Access data may help generate hypotheses about treatment exposure and patient subgroups, potentially informing future clinical research.
3. Dosing and safety patterns became easier to identify
The research also examined treatment modifications associated with low platelet and neutrophil counts.
Among the 39 treated patients:
15 patients (38%) experienced low blood counts that led to dosing changes or treatment delays.
10 patients required dose reductions, including eight whose reductions occurred during cycle two.
Five additional patients experienced treatment delays.
Patients who began treatment within 90 days of prior chemotherapy or radiotherapy showed a higher observed frequency of these modifications (11 of 21 versus 4 of 18), although the difference was not statistically significant (p = 0.098).
The analysis identified a recurring dose-reduction pattern, typically from 30 mg/m² to 20 mg/m².
These findings suggest questions for future investigation, including whether different starting doses might affect treatment exposure and tolerability.
The poster also noted that 12 of the 39 patients received VAL-083 alongside other therapies, including bevacizumab, Optune, or pembrolizumab.
Because this was a single-arm Expanded Access cohort, observations about longer treatment duration cannot be attributed to VAL-083 alone.
The significance: AI can make dosing and safety patterns easier to examine across patients, potentially helping clinical teams identify questions for prospective study and improve program oversight.
4. Independent AI analyses produced consistent results
Perhaps the most significant technology finding was the reproducibility of the analysis.
Researchers conducted four independent AI reviews using two models across two analytical environments.
All four analyses reproduced the same core cohort figures, including patient counts, demographic summaries, dose reductions, and recorded deaths.
Independent reviews also identified details missed during earlier analyses and helped correct errors.
Where interpretations differed, disagreements generally involved how particular clinical events should be categorized or counted rather than the underlying arithmetic.
The study also compared AI-generated results with two previously prepared human reports. Those reports differed on several patient-level details and a headline figure, while the AI analyses agreed on the core cohort figures.
This does not mean AI eliminates the need for human review.
In fact, the researchers emphasized that reliability depends on a disciplined process involving independent analyses, predefined counting rules, and verification against original records.
The significance: When supported by source-level traceability and appropriate validation, AI-assisted analysis can improve the consistency and reproducibility of real-world data reviews.
5. The research exposed opportunities to improve data collection
Another important finding was not about the investigational therapy itself, but about the quality and structure of Expanded Access data.
Critical information about disease progression, treatment response, adverse events, and patient outcomes frequently existed only in free-text notes.
For example, only 17 of the 39 treated patients had any adverse events recorded in the available records, although all serious adverse events were reported. Vital status was not recorded for 33 patients, while six deaths had documented dates.
These limitations made certain analyses difficult and highlighted the importance of distinguishing missing information from a confirmed absence of an event.
The research demonstrates that AI can help retrieve and organize unstructured information, but it cannot compensate for data that was never captured.
The significance: Pharmaceutical companies have an opportunity to improve the value of Expanded Access data by combining AI-powered analysis with more consistent, structured data collection.
From Individual Patient Records to a More Complete Clinical Picture
One particularly compelling example in the poster involved a patient with diffuse intrinsic pontine glioma (DIPG), an aggressive brain tumor.
The AI agent reconstructed a longitudinal patient history from 17 linked MedaSystems platform forms.
The resulting analysis brought together prior treatments, biomarker information, dosing history, treatment interruptions, and clinical outcomes.
The patient received 18 cycles of VAL-083 without a recorded dose reduction. Treatment was temporarily paused for re-irradiation and subsequently resumed for 13 additional cycles.
The reconstructed record documented progression-free survival of 13.8 months and overall survival of 16.9 months from the first VAL-083 dose.
This individual case cannot establish treatment efficacy. However, it demonstrates the ability of AI to assemble a comprehensive, traceable patient history from records that would otherwise require significant manual review.
For medical affairs and clinical research teams, this capability could make detailed case reviews more accessible and efficient.
The Bigger Opportunity: Making Expanded Access Data More Useful
The implications extend well beyond neuro-oncology.
Expanded Access Programs have historically been managed primarily as operational and regulatory activities. Pharmaceutical companies must coordinate requests, assess eligibility, manage approvals, track investigational product supply, monitor safety, and maintain appropriate documentation.
The data generated through these activities can also provide valuable insights into real-world treatment experience.
Yet extracting those insights has often required separate retrospective research efforts.
AI-powered technology offers an opportunity to bring operational program management and data analysis closer together.
The research demonstrated the potential to generate recurring reports directly from live program data, including information about:
Patient requests, approvals, and treatment status
Investigational product supply and accountability
Treatment exposure, dose modifications, and delays
Adverse events and serious adverse events
Available patient outcomes and data completeness
The researchers also developed a workflow for automatically generating weekly program status reports for clinical and program leadership.
Rather than manually compiling the same information repeatedly, teams could regenerate analyses from current records, verify patient counts, and flag missing data.
This creates the potential for a more continuous approach to program monitoring and real-world evidence generation.
Why This Matters for Pharmaceutical Companies
For pharmaceutical manufacturers, the research highlights three emerging opportunities.
More efficient program oversight. Medical Affairs, Clinical Operations, and Regulatory Affairs teams could gain faster visibility into patient activity, treatment patterns, and program performance without relying exclusively on manually compiled reports.
Better-informed clinical research. Patterns in treatment exposure, safety observations, and biomarker characteristics may help generate hypotheses for future studies, particularly in rare diseases and patient populations underrepresented in clinical trials.
Greater value from existing data. By making Expanded Access data easier to retrieve, analyze, and verify, organizations may be able to use information already collected through their programs more effectively.
Realizing these opportunities requires more than introducing an AI assistant.
It also requires secure data management, consistent documentation, appropriate access controls, auditability, and clear procedures for validating analytical outputs.
In a regulated life sciences environment, the ability to trace an AI-generated finding back to its underlying source is especially important.
How MedaSystems Is Advancing AI-Powered Expanded Access Management
MedaSystems is extending its platform beyond the operational management of Expanded Access Programs to help pharmaceutical companies unlock greater value from the real-world data those programs generate.
Through its MCP Server, currently in beta, compatible AI assistants can interact with program data using natural-language queries while respecting existing user permissions and maintaining source-level traceability.
Combined with the platform's centralized workflows, audit trails, and AI-powered patient health information redaction capabilities, this approach supports both effective program management and more accessible data analysis.
The SNO research provides a concrete demonstration of what becomes possible when AI is connected to real-world treatment records in a structured, auditable environment.
It also reinforces an important principle: the value of AI in life sciences is not simply how quickly it can generate an answer, but how reliably that answer can be verified.
Looking Ahead: From Access to Evidence
Expanded Access Programs will continue to play an essential role in helping patients with serious diseases access investigational treatments.
But as pharmaceutical companies invest in real-world evidence strategies, these programs may also become a more valuable source of clinical insights.
The findings presented at SNO demonstrate how AI can help bridge the gap between collecting real-world treatment data and making that information useful.
The next opportunity is to build on these capabilities through better structured data collection, repeatable analytical workflows, and continued evaluation of AI-generated findings.
For life sciences organizations, this represents an important evolution: Expanded Access Programs can support not only patient access and responsible program oversight, but also the generation of new research questions and a deeper understanding of treatment experience.
The future of Expanded Access is not just about managing access more efficiently. It is also about learning more from the patients and programs already being served.
Learn More
Read the MedaSystems press release announcing the findings presented at the 2026 Society for Neuro-Oncology Annual Meeting.
To learn how MedaSystems helps pharmaceutical companies manage Expanded Access Programs and analyze real-world program data, request a personalized demonstration.