Large language models show variable and context-dependent accuracy when staging and grading periodontitis from clinical text, according to a scoping review evaluating early evidence on artificial intelligence in periodontal diagnosis. Researchers synthesized findings from six original studies published between 2024 and 2026 to assess how well generative models apply the diagnostic framework established in the 2018 classification of periodontitis. Across the evaluated literature, overall diagnostic accuracy demonstrated wide variability, and the underlying evidence base does not support deploying existing models as autonomous clinical diagnostic tools.
The scoping review was conducted following the PRISMA-ScR reporting guidelines and a predefined protocol registered on the Open Science Framework. To map the current landscape of research, investigators searched PubMed, Web of Science, and Scopus on August 19, 2026, for English-language studies published from 2023 onward, supplementing database results with citation tracking. Eligible investigations were required to be original studies directly comparing diagnostic stages or grades generated by large language models against established clinical reference standards, with descriptive methodological quality appraised using the QUADAS-2 tool.
In total, the review identified six studies that met the eligibility criteria, representing a collective sample of 400 patients, clinical cases, or simulated patient vignettes. All included investigations evaluated models from the OpenAI GPT series and provided clinical data to the systems exclusively in text format. The types of case data varied substantially across the literature: two studies utilized real patient records, two employed published or author-developed clinical cases, and two relied on simulated patient vignettes.
Diagnostic performance varied broadly depending on the classification task and the nature of the clinical input. Staging accuracy ranged from 36.2 percent to 70.5 percent across the analyzed datasets, while grading accuracy spanned from 50.5 percent to 100 percent. Despite occasional high accuracy figures, statistical agreement between model determinations and reference standards remained slight to moderate, with reported Cohen’s kappa values falling between 0.18 and 0.48.
The review observed that system performance depended heavily on clinical context and presentation complexity. The single model evaluated across both simulated and authentic settings achieved markedly better results when processing clear-cut synthetic vignettes compared to authentic patient records that featured incomplete clinical inputs. Diagnostic errors primarily clustered between adjacent disease categories, though models exhibited inconsistent behavior regarding whether they systematically underestimated or overestimated the severity of periodontal breakdown.
The authors noted that comparative evaluations favoring alternative computational frameworks were limited. Separate individual studies suggested potential benefits for a task-specific neural network or retrieval-augmented platforms; however, each comparison stemmed from a single investigation accompanied by important methodological caveats. Furthermore, in the study examining retrieval-augmented platforms, the performance differences between competing systems were smallest specifically for periodontitis staging and grading tasks.
Methodological appraisal revealed important limitations across the available literature, as the risk of bias concerning patient selection was judged to be high or unclear in every analyzed study. The authors concluded that current evidence remains limited, recent, and heterogeneous, precluding autonomous diagnostic use. Future investigations require consecutive real-world patient cohorts with complete clinical and radiographic findings, independent reference standards, transparent reporting of model configurations, and direct evaluation of multimodal models capable of interpreting radiographs.
Methodological rigor in the scoping review was supported by having two reviewers independently select eligible studies and chart relevant data before the findings were synthesised narratively. This investigation addressed an important clinical question, as large language models are increasingly explored as clinical decision-support tools, yet their specific ability to apply the 2018 classification of periodontitis has remained unclear. To clarify the current evidence base, the review comprehensively mapped model performance across staging and grading tasks, the evaluation methods used by researchers, and the limitations reported across published studies.
By systematically charting these evaluation methods alongside reported diagnostic limitations, the review provided necessary context for the modest to moderate accuracy observed across textual data. Documenting how investigators designed their assessments and reference comparisons clarifies why current evidence remains heterogeneous, highlighting the exact methodological gaps that future periodontal research must address.
Subject of Research: Scoping review of large language models for periodontitis staging and grading
Article Title: Large language models for periodontitis staging and grading according to the 2018 classification: a scoping review
Article References: Karadag, I., & Tezel, A. (2026). Large language models for periodontitis staging and grading according to the 2018 classification: a scoping review. BMC Oral Health. https://doi.org/10.1186/s12903-026-10429-9
Image Credits: AI Generated
DOI: 10.1186/s12903-026-10429-9
Keywords: periodontitis, staging and grading, large language models, artificial intelligence, clinical decision support, scoping review, dental informatics, diagnostic accuracy, periodontal classification, OpenAI GPT, PRISMA-ScR, QUADAS-2, oral health, clinical vignettes, electronic health records, retrieval-augmented generation
News Source: Ophelia Keating. (November 20, 2026). Large Language Models Show Inconsistent Accuracy in Staging and Grading Periodontitis. Scienmag.



