TAVR-VLM: AUROC 0.896, 8.1% Hallucination Rate for TAVR planning

TAVR-VLM, a multimodal model for Transcatheter Aortic Valve Replacement planning, appeared on arXiv on 25 Jun 2026. The paper introduces Risk-Conditioned Causal Grounding Attention (R-CGA) and evaluates the model on M^3TAVR, a 1,482-patient cohort, reporting an AUROC of 0.896.

What is TAVR-VLM and how does it work?

TAVR-VLM is a framework that enforces an internal "Risk → Region → Word" grounding pathway using Risk-Conditioned Causal Grounding Attention (R-CGA), which compresses multimodal inputs into a causal risk bottleneck and produces a global risk mask. The model then constrains token-level generation with a support-projected causal consistency objective so generated text remains grounded within the risk-defined support mask.

The paper describes R-CGA as purifying dense visual features into a global risk mask, forming a model-internal causal risk bottleneck. During autoregressive generation, the support-projected causal consistency objective is applied so token selection is restricted to regions indicated by the risk mask, reducing diagnostic hallucinations where text lacks anatomical grounding.

How did TAVR-VLM perform on M^3TAVR?

On the M^3TAVR dataset of 1,482 patients, TAVR-VLM achieved an AUROC of 0.896, a CIDEr score of 0.936, and a hallucination rate of 8.1%. Those figures are presented in the paper as the new state-of-the-art for this task.

The evaluation emphasizes both predictive discrimination and report-generation quality: AUROC quantifies model discrimination, CIDEr measures text similarity for generated reports, and the paper highlights a drastic reduction in hallucination rate to 8.1 percent as a key outcome. The authors position these metrics as improvements in interpretability for evidence-based surgical AI.

Why it matters

Clinical planning for Transcatheter Aortic Valve Replacement depends on precise, anatomically grounded multimodal interpretation. By structuring an explicit causal pathway from risk estimation to region masking to word generation, TAVR-VLM tackles a core failure mode of multimodal models in medicine: diagnostic hallucination. If the reported AUROC, CIDEr and reduced hallucination rate hold up under external validation, clinicians would get more trustworthy, evidence-linked suggestions from an AI assistant rather than unconstrained free-text output.

The paper frames the reduction in hallucination rate as directly improving interpretability, which matters because interpretability affects clinicians' willingness to rely on automated guidance in high-stakes surgical planning.

What to watch

Look for independent replication on external TAVR cohorts and for availability of the model, code or data links the paper references under its "Code, Data and Media" section. Confirmation that R-CGA generalizes beyond M^3TAVR or public releases of the authors' code and data would be the next concrete milestones to validate the claimed gains.

References and core facts drawn from the arXiv submission titled "TAVR-VLM: Risk-Conditioned Causal Grounding for Hallucination-Resistant Report Generation" (submitted 25 Jun 2026), which reports evaluation on M^3TAVR (1,482 patients), AUROC 0.896, CIDEr 0.936, and a hallucination rate of 8.1%.