RESUMO
Subjectivity and ambiguity of visual field classification limits the accuracy and reliability of glaucoma diagnosis, prognostication, and management decisions. Standardised rules for classifying glaucomatous visual field defects exist, but these are labour-intensive and therefore impractical for day-to-day clinical work. Here a web-application, Glaucoma Field Defect Classifier (GFDC), for automatic application of Hodapp-Parrish-Anderson, is presented and validated in a cross-sectional study. GFDC exhibits perfect accuracy in classifying mild, moderate, and severe glaucomatous field defects. GFDC may thereby improve the accuracy and fairness of clinical decision-making in glaucoma. The application and its source code are freely hosted online for clinicians and researchers to use with glaucoma patients.
RESUMO
Large language models (LLMs) underlie remarkable recent advanced in natural language processing, and they are beginning to be applied in clinical contexts. We aimed to evaluate the clinical potential of state-of-the-art LLMs in ophthalmology using a more robust benchmark than raw examination scores. We trialled GPT-3.5 and GPT-4 on 347 ophthalmology questions before GPT-3.5, GPT-4, PaLM 2, LLaMA, expert ophthalmologists, and doctors in training were trialled on a mock examination of 87 questions. Performance was analysed with respect to question subject and type (first order recall and higher order reasoning). Masked ophthalmologists graded the accuracy, relevance, and overall preference of GPT-3.5 and GPT-4 responses to the same questions. The performance of GPT-4 (69%) was superior to GPT-3.5 (48%), LLaMA (32%), and PaLM 2 (56%). GPT-4 compared favourably with expert ophthalmologists (median 76%, range 64-90%), ophthalmology trainees (median 59%, range 57-63%), and unspecialised junior doctors (median 43%, range 41-44%). Low agreement between LLMs and doctors reflected idiosyncratic differences in knowledge and reasoning with overall consistency across subjects and types (p>0.05). All ophthalmologists preferred GPT-4 responses over GPT-3.5 and rated the accuracy and relevance of GPT-4 as higher (p<0.05). LLMs are approaching expert-level knowledge and reasoning skills in ophthalmology. In view of the comparable or superior performance to trainee-grade ophthalmologists and unspecialised junior doctors, state-of-the-art LLMs such as GPT-4 may provide useful medical advice and assistance where access to expert ophthalmologists is limited. Clinical benchmarks provide useful assays of LLM capabilities in healthcare before clinical trials can be designed and conducted.
RESUMO
BACKGROUND: Large language models exhibiting human-level performance in specialized tasks are emerging; examples include Generative Pretrained Transformer 3.5, which underlies the processing of ChatGPT. Rigorous trials are required to understand the capabilities of emerging technology, so that innovation can be directed to benefit patients and practitioners. OBJECTIVE: Here, we evaluated the strengths and weaknesses of ChatGPT in primary care using the Membership of the Royal College of General Practitioners Applied Knowledge Test (AKT) as a medium. METHODS: AKT questions were sourced from a web-based question bank and 2 AKT practice papers. In total, 674 unique AKT questions were inputted to ChatGPT, with the model's answers recorded and compared to correct answers provided by the Royal College of General Practitioners. Each question was inputted twice in separate ChatGPT sessions, with answers on repeated trials compared to gauge consistency. Subject difficulty was gauged by referring to examiners' reports from 2018 to 2022. Novel explanations from ChatGPT-defined as information provided that was not inputted within the question or multiple answer choices-were recorded. Performance was analyzed with respect to subject, difficulty, question source, and novel model outputs to explore ChatGPT's strengths and weaknesses. RESULTS: Average overall performance of ChatGPT was 60.17%, which is below the mean passing mark in the last 2 years (70.42%). Accuracy differed between sources (P=.04 and .06). ChatGPT's performance varied with subject category (P=.02 and .02), but variation did not correlate with difficulty (Spearman ρ=-0.241 and -0.238; P=.19 and .20). The proclivity of ChatGPT to provide novel explanations did not affect accuracy (P>.99 and .23). CONCLUSIONS: Large language models are approaching human expert-level performance, although further development is required to match the performance of qualified primary care physicians in the AKT. Validated high-performance models may serve as assistants or autonomous clinical tools to ameliorate the general practice workforce crisis.