Application of Large Language Models in the Recommendation of Treatment Plans for Gastrointestinal Cancers

The medical records of over 1,200 patients with gastrointestinal cancers will be collected retrospectively from participating hospitals. This data will be split into an exploratory dataset (n=200) and a validation dataset (n>=1,000). Within the exploratory dataset, various prompt methods will be used to determine the treatment plans suggested by ChatGPT. Additionally, several clinicians of varied seniority levels will provide their treatment recommendations. For the validation dataset, ChatGPT's suggestions for treatment plans will undergo both qualitative and quantitative assessments by a multidisciplinary consultation (MDT) team. The recommendations from ChatGPT will then be compared with those from the clinicians.

Furthermore, this study will incorporate a prospective dataset comprising 400 participants with gastrointestinal cancers. The participants will be randomly allocated to either a control group (n=200) or a ChatGPT-Assisted group (n=200). In the control group, treatment plan recommendations will solely be provided by the clinicians and will guide subsequent treatments. In the ChatGPT-Assisted group, initial treatment plan recommendations will be independently proposed by both ChatGPT and the clinicians. Based on ChatGPT's suggestions, clinicians might selectively adjust their initial plans. Participants will then receive treatments as per these refined plans. Within the ChatGPT-Assisted group, the treatment plans of the initial 100 participants will be evaluated to determine the percentage of patients whose treatment plans are influenced by ChatGPT. Subsequently, the proportion of participants in the entire ChatGPT-Assisted group with treatment plans modified by ChatGPT will be calculated. The study will further monitor the 3-year progression-free survival (PFS) and the 5-year overall survival (OS) rates, contrasting the outcomes between the control and ChatGPT-assisted groups.