PanVasc Research for AI assisted evidence analysis in panvascular intervention

PanVasc Research introduces an executable framework for conducting vascular intervention studies, preserving the integrity of the original source data. Utilizing a fixed Qwen3-4B model, the study evaluated two key tasks using 50 ClinicalTrials.gov records from five different vascular query groups. The records were split into tests preserving the original source data, removing the source, primary outcome, or timeframe, and 50 clean controls.

Additionally, a separate 100-statement sample from the NLI4CT test set tested clinical-trial entailment and evidence selection accuracy against existing expert labels.

The study utilized both generic and checklist prompts, ensuring identical evidence and maximum generation budgets for each output. The outputs were also evaluated using an input-only deterministic contract. The registry exact accuracy was found to be 25.5% with the generic prompt and 38.5% with the checklist. The intact-case agreement was 50% with both prompts. The rule baseline recovered all 200 tuples, while the NLI label accuracy was 48% and 50% for the two arms, respectively.

The source contract retained 35 and 31 incorrect NLI labels in the two arms. The framework maintains provenance-recorded literature retrieval, structured research planning, and local numerical analysis. The experiments support a bounded assessment of source handling and semantic failure, rather than introducing a new foundation model or autonomous scientific discovery. The proposed endpoint representation highlights the need for additional domain annotation and independent validation for a panvascular research model.