Diagnostic accuracy and repeatability of ChatGPT using textual and radiographic data in reversible pulpitis: a retrospective diagnostic study.

OBJECTIVES

Large language models have recently evolved into multimodal systems capable of interpreting images alongside text. The aim of this study was to assess the diagnostic accuracy and repeatability of ChatGPT-5 for reversible pulpitis, based on structured clinical information and periapical radiographs (PRs).

METHODS

Thirty expert-validated cases of reversible pulpitis were retrospectively selected. For each case, clinical data and PRs were compiled to create a reference dataset, with expert consensus serving as the reference standard. A structured prompt incorporating clinical information and PR was designed and included six predefined questions: four addressing diagnostic reasoning, one assessing whether PR information was used to establish the diagnosis, and one requiring radiographic findings description. Each case was entered thirty times into ChatGPT-5, generating a total of 900 answers. Answers were assessed by two independent experts, with disagreements resolved by a third. Radiographic descriptions were graded using a three-point Likert scale. Diagnostic accuracy and repeatability were analysed using binomial and agreement methods.

RESULTS

Diagnostic questions related to tooth identification, pulpal diagnosis, periapical diagnosis and treatment indication showed accuracies above 96%. Overall accuracy for radiographic description was low (4.67%). Coronal radiographic interpretation showed poor performance (5.15%), whereas accuracy for pulpal and periapical structures exceeded 96%. Repeatability was almost perfect for diagnostic questions and pulpal/periapical interpretation, but moderate for coronal descriptions.

DISCUSSION: ChatGPT-5 demonstrated high accuracy and repeatability in text-based diagnostic reasoning for reversible pulpitis. However, limitations in radiographic interpretation, particularly of coronal structures, indicate that its use should be limited to a supervised decision-support role.