Automated survivorship care plan generation using large language models.

Abstract

e13595 Background: Cancer survivorship, particularly from post-treatment to end-of-life, faces challenges such as communication gaps and coordination issues, leading to suboptimal care. A survivorship care plan (SCP) is a summarized record of patient cancer history and treatments, along with follow-up care recommendations. Despite extensive survivorship care guidelines from national cancer organizations, their adoption remains limited. Moreover, creating SCPs is time-intensive and burdensome for healthcare providers due to workforce limitations, hindering efficient personalization of care plans. Our study aims to address these issues by employing large language models (LLMs) to generate personalized SCPs. Methods: We introduce SurvGPT, a customized LLM designed to generate personalized SCPs focusing on its two main sections: the treatment summary and the follow-up care plan. Initially, our method extracts information from the structured and unstructured patient records to generate the treatment summary. Subsequently, it retrieves relevant context from a survivorship knowledge base crafted from existing medical literature and guidelines to generate a customized follow-up care plan. To mitigate LLM-related challenges such as hallucinations, our method ensures that the generated follow-up care plan is grounded on existing medical knowledge with a self-checking mechanism. Results: We created synthetic patient data based upon clinical inputs by a medical oncologist and an internist. We conducted a preliminary study using synthetic lung cancer patient EHR data to evaluate our approach. The evaluation focused on several key dimensions to validate the quality of the generated SCPs, including comprehensiveness, personalization, clarity, and actionable recommendations. The results presented in the table suggest promise for this approach yet also need for additional refinements both in degree of details extracted for ongoing care and then in formatting for effective use. The results also illustrate the complexity of generating synthetic cases for this work whose similarity to real patient care improved with iteration yet remained on occasion perceptibly different from current practice as noted by our clinical oncologist team member. Conclusions: In conclusion, our study demonstrates the potential of SurvGPT for generating personalized SCPs. The results of the preliminary results suggest that this approach is technically feasible with further validation required using real patient data. [Table: see text]