Diagnostic interpretation of corneal tomography using a multimodal large language model (ChatGPT)

Résumé fourni par la source

Purpose: To describe the use of a commercially available general large language model (LLM) in extracting and interpreting several key metrics from entire raw corneal tomography (Pentacam) reports for the diagnosis of corneal disorders. Observation: Anonymized corneal tomography biometry reports of 50 eyes of 50 patients with healthy corneas (n = 28), keratoconus (n = 20) and post-surgical ectasia (n = 2) were analyzed by a multimodal general LLM. System prompts were used to extract flat and steep keratometry values (K1 and K2, respectively), astigmatism, pachymetry values and provide an overall diagnosis. Accuracy of data extraction was 100 % across all metrics and the model provided a diagnosis in agreement with the two observers in all eyes. Pachymetry and maximum keratometry values were the most common metric used to formulate the diagnosis and was cited in all eyes. This was followed by specifically citing the highest elevation map values (88 %) and degree of astigmatism (74 %). Conclusion and importance: In this proof-of-concept study, a commercially available multimodal LLM was able to extract data from raw corneal tomography reports with high accuracy and with retention of spatial context, and formulated correct diagnoses with excellent proficiency. This study demonstrates the use of emerging LLMs as diagnostic adjuncts through the synthesis of multimodal data.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.