Multimodal LLM periodontitis risk assessment and prevention planning from EHR data and panoramic radiographs

Source article: Periodontitis Risk Assessment and Prevention Planning: Comparative Study of Multimodal Large Language Models and Periodontist Evaluations

Background Periodontitis is one of the most prevalent yet preventable oral diseases, as indicated by multiple clinical and radiographic factors. As these factors are recorded in electronic health records (EHRs), their reuse offers opportunities for personalized risk assessment and targeted prevention. Predictive AI and traditional machine learning models support fragmented detection tasks but lack the integration of textual and imaging predictors. Emerging multimodal large language models (M-LLMs) show promise i…

By the Truvace deskOctober 14, 20262 min read1 sourcePeer-reviewedTRV-2026-1371· cite & verify

Periodontitis Risk Assessment and Prevention Planning: Comparative Study of Multimodal Large Language Models and Periodontist Evaluations — "09-8018-23" by NavyMedicine is marked with Public Domain Mark 1.0. To view the terms, visit https://creativecommons.org/publicdomain/mark/1.0/.

Trace impact readingContested

P 69

Both sides are scored from claims and sources, not community votes.

G 65

In brief

Researchers tested three reasoning-capable multimodal language models on ten periodontal vignettes, each combining a panoramic radiograph with electronic health record text. Periodontal experts rated the models' outputs, and a periodontist's manually produced output on the same prompts served as the reference.

The results suggest these tools could one day act as digital assistants for prevention planning, but only if they reach periodontist-level quality. Radiograph interpretation remains the weak point, and it is unclear whether better prompting or newer model versions will fix it.

Main points

Ten vignettes, each pairing a panoramic radiograph with textual EHR data, were used, following TRIPOD guidance for evaluating LLMs.
Experts scored outputs on 6 predefined criteria using a 5-point Likert scale.
GPT o1 Pro and Claude Sonnet 4 reached 86.7% and 85.6% acceptable ratings versus 87.8% for the periodontist.
Gemini 2.5 Pro reached only 59.4% acceptable ratings and was rated significantly lower than the others.
Outputs were still rated well when radiographic findings were wrong, suggesting EHR data alone give a substantial basis.

The gain

Two of three multimodal language models produced periodontitis risk assessments and prevention plans that experts rated acceptable at rates comparable to a periodontist's.

The problem

All tested models made unacceptable errors in interpreting panoramic radiographs, and one model scored markedly lower than the periodontist and the other models.

The rundown

The study was motivated by the fact that periodontitis is common but preventable, and that clinical and radiographic factors recorded in EHRs could support personalized risk assessment. Earlier predictive AI and machine learning tools addressed fragmented detection tasks and did not integrate text with imaging.

Three models able to reason and handle multimodal data were compared with a periodontist working from the same prompts and inputs. Statistical analyses examined overall performance per model and whether it varied by model, scenario complexity, or rater.

The authors conclude that the models showed strong reasoning in periodontal assessment, and that unacceptable elements across all models traced to radiographic misreads. They state that clinical use requires at least periodontist-comparable quality, a bar that GPT and Claude appear to approach.

What this doesn’t fix

The study used only ten vignettes and three models, and radiographic interpretation errors occurred across all of them; the authors say improvement from refined prompting or newer versions is only possible, not shown.