Abstract B008: Fine-tuned large language model accurately estimates progression-free survival from radiology reports in immune checkpoint inhibitor–treated cancers

Abstract Introduction: The development of predictive models for cancer treatment outcome prediction depends on datasets with accurately curated clinical outcomes, such as progression-free survival (PFS). However, current approaches to determining real-world PFS require extensive manual review of imaging reports and clinical documentation—a time- and labor-intensive process prone to variability. In this study, we evaluated whether a fine-tuned LLM could accurately estimate PFS using only longitudinal radiology reports from patients treated with immune checkpoint inhibitors (ICIs). Methods: A previously annotated dataset of 972 patients with 7 different advanced solid tumor cancer types (non-small cell lung, melanoma, renal cell, head and neck squamous cell, gastroesophageal, bladder, and breast) treated with immune checkpoint blockade at Stanford University were included in this study. A Gemini Flash (version 1.5) model instance was fine-tuned for PFS prediction within a secure Google Cloud Platform environment. Patients were used for model fine-tuning or validation with a 70:30 tuning/validation split stratified by tumor type. The model was fine-tuned with all available advanced modality imaging reports (eg, CT, PET-CT, and MRI) from the date of treatment initiation up to the first report describing disease progression for each patient in tuning dataset. To evaluate the model in the validation dataset, each available report following treatment initiation was presented to the model in sequence until progression was determined to have occurred or the patient was censored. Results: There was no significant difference between the actual PFS and the PFS predictions from the fine-tuned model in the validation dataset (median PFS by Kaplan-Meier analysis 8.07 months vs. 7.28 months, p=0.136 by log-rank test). The concordance index (C-index) between the fine-tuned model predictions and actual PFS was 0.756. Zero-shot deployment of the same Gemini model resulted in a significant difference in actual and predicted PFS (median PFS 8.07 months vs. 2.72 months p < 0.0001) and C-index of 0.712. Conclusions: A fine-tuned large language model can accurately identify disease progression in cancer imaging reports and determine progression-free survival in response to treatment with immune checkpoint blockade. Attempting this task without fine-tuning results in similar estimations of relative progression risk within the dataset, but systemic underestimation of absolute progression-free survival times. Supervised fine-tuning is valuable for performance optimization on complex cancer outcome curation tasks by AI models. Citation Format: Feyisope Eweje, Riya Dulepet, Yijiang Chen, Ruijiang Li. Fine-tuned large language model accurately estimates progression-free survival from radiology reports in immune checkpoint inhibitor–treated cancers [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B008.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.