Digital health · 2026
Head-to-head evaluation of ChatGPT, DeepSeek, and Perplexity on acid-base disorder case clinical management and drug treatment: Accuracy, domain performance, and response consistency assessment.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Khobrani Moteb, Asaad Khalil Asaad Ahmed, Ahmad Salman Ashfaq, Ishaqui Azfar Athar
Аннотация
BACKGROUND: Large language models (LLMs) are increasingly used in medical education, but their performance and reliability on mechanistically demanding topics like acid-base interpretation are unclear.
METHODS: We conducted a cross-sectional head-to-head evaluation of three LLMs (ChatGPT, DeepSeek, Perplexity) using 75 textbook acid-base cases related to management and drug treatment. One obsolete case was excluded, leaving 74 vignettes (29 metabolic, 10 respiratory, 35 mixed) that generated 510 multiple-choice questions (MCQs). Each MCQ was mapped to case category, one of seven cognitive domains, and one of four distractor-based difficulty bands. All questions were posed via public web interfaces in two independent phases (Phase I and II). Primary outcome was accuracy (proportion correct); consistency outcomes were identical answers across phases and reproducibly correct answers (correct in both phases).
RESULTS: Overall accuracy was 59.8% (305/510) for ChatGPT, 58.8% (300/510) for DeepSeek, and 53.7% (274/510) for Perplexity (Cochran's Q p = 0.16); ChatGPT and DeepSeek each outperformed Perplexity (p = 0.0086 and p = 0.0374). In metabolic cases (164 items), Perplexity scored 50.6% versus 60.4% and 62.8% for ChatGPT and DeepSeek. Accuracy was highest for compensation/expected response (116 items; 70.7% for ChatGPT and DeepSeek, 62.1% for Perplexity) and lowest for metabolic acidosis subtypes (59 items; 52.5%, 42.4%, 32.2%). Accuracy declined with difficulty, with Perplexity significantly lower than the other models in moderate-hard and hard bands. Identical answers across phases occurred in 81.7% of items for ChatGPT, 81.1% for DeepSeek, and 76.1% for Perplexity, but reproducibly correct answers were 52.1%, 48.8%, and 43.2%.
CONCLUSIONS: ChatGPT and DeepSeek showed moderate performance and generally higher accuracy than Perplexity. Stable repeated errors despite only moderate reproducibly correct responses support cautious interpretation of LLM outputs and argue against their use as stand-alone tools for acid-base learning or clinical decision making of management and drug treatment .
METHODS: We conducted a cross-sectional head-to-head evaluation of three LLMs (ChatGPT, DeepSeek, Perplexity) using 75 textbook acid-base cases related to management and drug treatment. One obsolete case was excluded, leaving 74 vignettes (29 metabolic, 10 respiratory, 35 mixed) that generated 510 multiple-choice questions (MCQs). Each MCQ was mapped to case category, one of seven cognitive domains, and one of four distractor-based difficulty bands. All questions were posed via public web interfaces in two independent phases (Phase I and II). Primary outcome was accuracy (proportion correct); consistency outcomes were identical answers across phases and reproducibly correct answers (correct in both phases).
RESULTS: Overall accuracy was 59.8% (305/510) for ChatGPT, 58.8% (300/510) for DeepSeek, and 53.7% (274/510) for Perplexity (Cochran's Q p = 0.16); ChatGPT and DeepSeek each outperformed Perplexity (p = 0.0086 and p = 0.0374). In metabolic cases (164 items), Perplexity scored 50.6% versus 60.4% and 62.8% for ChatGPT and DeepSeek. Accuracy was highest for compensation/expected response (116 items; 70.7% for ChatGPT and DeepSeek, 62.1% for Perplexity) and lowest for metabolic acidosis subtypes (59 items; 52.5%, 42.4%, 32.2%). Accuracy declined with difficulty, with Perplexity significantly lower than the other models in moderate-hard and hard bands. Identical answers across phases occurred in 81.7% of items for ChatGPT, 81.1% for DeepSeek, and 76.1% for Perplexity, but reproducibly correct answers were 52.1%, 48.8%, and 43.2%.
CONCLUSIONS: ChatGPT and DeepSeek showed moderate performance and generally higher accuracy than Perplexity. Stable repeated errors despite only moderate reproducibly correct responses support cautious interpretation of LLM outputs and argue against their use as stand-alone tools for acid-base learning or clinical decision making of management and drug treatment .