How well it works: Benchmarking performance of GPT models on medical natural language processing tasks

Abstract

Abstract Importance The ability of large language models (LLMs) to generate high-quality, human-like text has been accompanied with speculation about their application in healthcare, alongside ethical and safety concerns. Objective Evaluate LLM performance on medical natural language processing (NLP) tasks, benchmarked against other commercially available tools. Design Observational study to evaluate and compare model performance. All models were commercially available and were evaluated without modification. Setting The Text Analysis Coding (TAC) 2017 challenge was used to assess ability to perform medical coding using standard MedDRA preferred terms. Text from 55 publicly available de-identified medical transcription reports were annotated to identify pre-defined medical concepts (age, disease/symptom, body structure, medication name, and medication dosage). Participants Publicly available, de-identified adverse event and medical transcription reports were used for evaluation. Exposures For each task, general LLMs (GPT-3.5-turbo, GPT-4) were compared to commercially available healthcare NLP tools (Microsoft Text Analytics for Health, Amazon Comprehend Medical, IQVIA API Marketplace). Main Outcomes and Measures For each NLP task, sensitivity, positive predictive value (PPV) and F1 score were calculated. Because GPT models had variable outputs, the range of metrics over 5 trials is reported. Results For MedDRA coding, GPT-4 had similar F1 score performance to healthcare NLP algorithms (GPT-4: 0.67 to 0.73; Microsoft Text Analytics for Health: 0.66, IQVIA API Marketplace: 0.72), while GPT-3.5-turbo had considerably lower performance (0.50 to 0.51). For medical information extraction, LLM performance varied widely across differing medical concepts; the highest F1 scores were for age (GPT-3.5-turbo: 0.82 to 0.83, GPT-4: 0.84 to 0.87) and medication name (GPT-3.5-turbo: 0.55 to 0.59, GPT-4: 0.70 to 0.76), while F1 scores for disease/symptom, body structure, and medication dosage were lower than those observed for the healthcare NLP tools. GPT-3.5-turbo and GPT-4 generally had lower sensitivity than comparators. Conclusions and Relevance In the absence of domain-specific fine tuning, GPT-4 performed similarly to healthcare-specific NLP tools on some tasks and less accurately on others; GPT-3.5-turbo was consistently less accurate than comparators. To maximize benefit and reduce risk of harm, robust quantitative evaluation for specific tasks should be performed prior to implementing LLMs in medical contexts.

Ähnliche Arbeiten