P-OGC57 Predicting survival and response to therapy using diagnostic biopsies: A machine learning approach to facilitate treatment decisions for oesophageal adenocarcinoma

Abstract Background Standard of care for locally advanced oesophageal adenocarcinoma is neoadjuvant chemotherapy or chemoradiotherapy followed by surgery. Only a minority patients (<25%) derive significant survival benefit from treatment and there are no reliable means establishing prior to in whom this will occur. Moreover, accurate prediction also not possible. The availability machine learning techniques provides the potential use complex data sources answer these problems. In study, we assessed utility high-resolution digital microscopy pre-treatment biopsies predicting both response therapy overall survival. Methods A total 157 cases were included study. Pre-treatment clinical information, including treatment, was obtained, along with diagnostic biopsies. Diagnostic converted into whole slide-images features extracted using pre-trained convolutional neural network Xception. Single representative images patches which predictive models trained. Elastic net regression classifiers derived validated bootstrapping 1000 resampled datasets. considered according Mandard tumour grade (TRG). Model performance quantified C-index (for TRG) time-dependent AUC (tAUC, fo Overall survival) calibration plots. Results Median 78.9months (95%CI 35.9 months – reached). Survival at 5-years 52.1%. Neoadjuvant received 123 (78.3%), seen 45 (36.6%). more likely those who than (53.3% vs 23.1% p < 0.001) older (median age 69.4 66.0 years, = 0.038), other characteristics similar. model image achieved good discrimination (C-index 0.767, 95%CI 0.701-0.833) calibration. Accuracy modest (tAUC 0.640, 0.518-0.762). Conclusions Using small dataset, feature extraction pipeline patient level outcomes has been demonstrated. This marked survival, may reflect importance determining former outcome. Further study refine methodology confirmation larger datasets required before expansion settings.