Use of a computational histology artificial intelligence-powered predictive biomarker for chemotherapy selection in advanced pancreatic cancer patients from a multi-institutional cohort including two prospective studies.

Abstract

764 Background: This study used the previously developed Computational Histology Artificial Intelligence (CHAI) platform to develop and validate a pathology-derived signature to distinguish patients with advanced pancreatic ductal adenocarcinoma (PDAC) likely to benefit from first-line fluoropyrimidine-based (F-chemo) versus gemcitabine-based (G-chemo) chemotherapy regimens. Methods: Whole slide images of H&E stained diagnostic biopsy specimens and clinical data were used. The development set was a real-world cohort of advanced PDAC patients treated with first-line F-chemo or G-chemo regimens from two academic medical centers. The CHAI platform was used to extract quantitative histomorphologic features, and then compose a continuous score associated with the primary endpoint of time to next treatment or death (TNTD), that was then dichotomized into a G-pref or F-pref result. The biomarker and threshold were locked. An independent validation cohort composed of patients from the prospective COMPASS trial and Know Your Tumor Registry was then used to assess the performance of the biomarker to predict improved TNTD and overall survival (OS) from fluoropyrimidine-based versus gemcitabine-based regimens. Results: The study cohort constituted 477 patients (178 in the development cohort, 299 in the validation cohort). In the validation cohort among the 173 F-pref patients, those treated with F-chemo had significantly better outcomes than G-chemo for both the TNTD (HR=0.68, p=0.036) and OS (HR = 0.57, p=0.003) endpoints. Among the 126 G-pref patients, those with G-chemo had significantly superior TNTD (HR=0.65, p=0.039), but no difference in OS (HR=0.87, p=0.6) compared to those with F-chemo. In multivariate cox proportional hazards models of TNTD and OS, the biomarker predicted differential treatment effect with significant biomarker-treatment interaction terms (TNTD: p=0.003; OS: p=0.016). Conclusions: The CHAI-powered signature developed from a multi-institutional real-world cohort and validated on a prospectively collected cohort predicted treatment efficacy, as measured by TNTD and OS, with fluoropyrimidine- versus gemcitabine-based chemotherapy. This biomarker can guide optimal treatment selection for first-line therapy in advanced PDAC. TNTD and OS in a validation cohort composed of data from two prospective studies, stratified by the biomarker. F-Chemo Median (95% CI) G-Chemo Median (95% CI) Cox Proportional Hazards Model Biomarker-Treatment Interaction Likelihood Ratio Test p-value TNTD p= 0.003 F-pref 8.6 (7.4-11.3) 7.5 (5.8-8.7) G-pref 7.2 (6.1-8.7) 9.6 (7.1-13.6) OS p= 0.016 F-pref 14.4 (11.3-16.7) 11.7 (7.8 - 12.7) G-pref 12.4 (11.1 - 14.5) 14.3 (9.0-21.3)