Early-phase clinical trials are the high-stakes trials in oncology, where decisions are made based on a few dozen patients and are key for drug development. When AI enters this domain, it enters an area that is not only high stakes but also high risk, according to Julien Vibert of Institut Gustave Roussy, who argues AI early-phase clinical trials oncology applications need to be judged on evidence rather than hype.
Why Patient Matching Is the First Test for AI Early-Phase Clinical Trials Oncology
Finding the right patients is the biggest challenge for early-phase clinical trials, since having no patients, or not having the right ones, can undermine the research entirely. Vibert notes large language models are well suited to this task, pointing to TrialGPT, a tool that automatically matches patients to trials and reduced screening time by 43% compared to manual matching, with performance comparable to human experts.
Why Real-World Validation Tempered This Early Promise
Despite TrialGPT’s strong initial results, Vibert cautions that real-world prospective validation later found its precision and recall were not as strong, illustrating a recurring theme in his assessment: promising retrospective results don’t always hold up once tools are tested prospectively in real clinical settings.
The FDA-Qualified Endpoint Within AI Early-Phase Clinical Trials Oncology
Another area, still less mature, is deriving AI-based endpoints directly. Vibert cites AIM-NASH, a deep learning and computational pathology tool that estimates liver disease severity in non-alcoholic steatohepatitis, which the FDA qualified in 2025 as an AI-derived endpoint usable in clinical trials to demonstrate drug efficacy.
Why Biomarker Validation Remains Non-Negotiable
Vibert frames this FDA qualification as a genuine proof of concept that AI-derived endpoints can help design trials and accelerate efficacy prediction, but stresses that biomarkers always need to be validated, referencing the ESMO Basic Requirements for AI-Based Biomarkers in Oncology as the governing standard for this validation process.
Detecting Adverse Events Within AI Early-Phase Clinical Trials Oncology
Large language models are particularly good at analyzing text, and in early-phase trials they can analyze clinical datasets to flag potential adverse events or safety signals that need detection at very early stages. Vibert points to Resilience, a spin-off from his own institution now deployed in multiple countries, an app collecting patient-reported outcomes with integrated AI analysis designed to detect adverse-event signals more quickly than researchers working manually.
Why Academic-Built Chatbots Matter Here
Vibert argues that while patients already use commercial chatbots, academics need to develop specialized chatbots that can genuinely improve how patients are monitored in early-phase trials, distinguishing purpose-built clinical monitoring tools from general-purpose consumer AI applications.
Digital Twins and Synthetic Data Within AI Early-Phase Clinical Trials Oncology
Digital twins and synthetic data represent another hyped area with potential importance for phase 1 trials, where researchers would ideally want a control arm but historical controls carry limitations. Digital twins, virtual models of patients, can simulate counterfactuals and the effects of different drugs to assess outcomes, functioning as a counterfactual model against which different treatments can be tested.
Why Vibert Calls These Proof-of-Concept Models Rather Than Mature Tools
Vibert cautions that despite the hype, these remain proof-of-concept models unlikely to generalize unless properly designed and trained on sufficient data, with risks of hallucination and bias always present, particularly with large language models.
Adaptive and Agentic Approaches Within AI Early-Phase Clinical Trials Oncology
Vibert describes emerging “prismatic” trial designs, a patient-centered approach where dosing and treatment allocation adapt based on a patient’s multimodal data, as well as ADAPT/ARPA-H, a U.S. blueprint for a learning cancer treatment system that adjusts trials according to tumor biology and evolution.
Why Agentic AI Represents the Next Conceptual Frontier
Vibert suggests agentic AI, systems capable of reasoning, planning, deciding, and taking action, could eventually help determine optimal treatment decisions based on multimodal data analyzed by AI, though he frames this as a future possibility rather than a currently validated capability.
What This AI Early-Phase Clinical Trials Oncology Assessment Means Going Forward
Vibert concludes that these are powerful tools worth counting on, but AI is not yet a co-investigator, and prospective validation remains key to the field’s progress. Given his consistent emphasis on the gap between retrospective promise and prospective validation, seen clearly in TrialGPT’s real-world performance decline, researchers and trial sponsors evaluating similar AI tools should prioritize prospective testing before relying on any tool’s initial published results.
What to Watch Going Forward
As tools like digital twins, adaptive prismatic trial designs, and agentic AI treatment decision-making continue developing, the oncology research community will likely watch for additional FDA-qualified AI endpoints following AIM-NASH’s precedent, and for prospective validation studies testing whether these conceptual approaches perform as well in practice as their early proof-of-concept results suggest. Given Vibert’s explicit caution about hallucination and bias risks in large language models, this AI early-phase clinical trials oncology landscape will likely continue evolving cautiously, with rigorous validation remaining the deciding factor for which tools move from hype into genuine clinical trial infrastructure.
For more healthcare industry updates, insights and news, visit DistilINFO. Click here to subscribe to stay informed.
