European journal of surgical oncology : the journal of the European Society of Surgical Oncology and the British Association of Surgical Oncology · 2026
Virtual patient modeling for generative-AI-assisted treatment decision-making in lymphedema care: AI tends to favor more aggressive treatment.
СтатьяНаблюдательное исследованиеНазвание 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.
Tsujimoto Yuki, Shiraishi Makoto, Yamanaka Hiroki, Lee Haesu, Okazaki Mutsumi, Morimoto Naoki
Аннотация
BACKGROUND: The rapid advancement of generative artificial intelligence (gen-AI) has prompted interest in whether it can recognize and respond to individual clinical backgrounds in treatment decision-making. To explore this, we developed a virtual patient model for lymphedema and conducted an observational study to examine what treatments AI would recommend, whether the recommendations were individualized, and what tendencies the AI exhibited.
METHODS: A virtual cohort of 100 patients with secondary upper extremity lymphedema following breast cancer surgery was constructed using generative pre-trained transformer-4 omni (GPT-4o). For each virtual patient, six clinical questions, based on the Japanese Lymphedema Guidelines 2024, were submitted to the AI to elicit individualized recommendations. The answers obtained were compared with guidelines-defined recommendation levels to assess concordance, deviation, and treatment tendencies, analyzed by patient factors.
RESULTS: Multivariate analysis demonstrated that GPT-4o-generated recommendations were tailored to individual patient characteristics. They showed high concordance with guideline-defined recommendations for conservative care but greater variability and bias toward invasive options in surgical contexts.
CONCLUSION: The preference of gen-AI for invasive treatments may reflect an overestimation of the benefits of performing treatments rather than withholding them, especially in invasive treatments. This bias shows a limitation of current gen-AI in complex decisions. Our reproducible simulation framework identified this bias and variability, clarifying both strengths and limitations of gen-AI. This type of AI-on-AI observational study may help improve the accuracy of AI and support its future role in clinical care.
METHODS: A virtual cohort of 100 patients with secondary upper extremity lymphedema following breast cancer surgery was constructed using generative pre-trained transformer-4 omni (GPT-4o). For each virtual patient, six clinical questions, based on the Japanese Lymphedema Guidelines 2024, were submitted to the AI to elicit individualized recommendations. The answers obtained were compared with guidelines-defined recommendation levels to assess concordance, deviation, and treatment tendencies, analyzed by patient factors.
RESULTS: Multivariate analysis demonstrated that GPT-4o-generated recommendations were tailored to individual patient characteristics. They showed high concordance with guideline-defined recommendations for conservative care but greater variability and bias toward invasive options in surgical contexts.
CONCLUSION: The preference of gen-AI for invasive treatments may reflect an overestimation of the benefits of performing treatments rather than withholding them, especially in invasive treatments. This bias shows a limitation of current gen-AI in complex decisions. Our reproducible simulation framework identified this bias and variability, clarifying both strengths and limitations of gen-AI. This type of AI-on-AI observational study may help improve the accuracy of AI and support its future role in clinical care.