Prediction of five-year survival of patients with esophageal cancer and the effect of biomarkers on predictive performance using Artificial Intelligence

Abstract

Abstract AIM We use Artificial Intelligence (AI) to predict the long-term survival of patients with resectable esophageal cancer (EC). We test whether AI can predict survival with all available data, with data at the time of primary staging (PS) and if biomarker analysis is equivalent to the Tumor, Node, Metastasis (TNM) classification in survival predictions. METHODS This retrospective study included 1002 patients with EC, 55 patient characteristics, and 55 biomarkers after surgical treatment between 1996 and 2021. The dataset was divided into short-term survival (overall survival, OS: > 90 days but < 5 years + recorded death) and long-term survival (OS: ≥ 5 years). AI methods (Random Forest RF, XG Boost XG, Artificial neural network ANN, TabNet TN) and logistic regression (LR) were used for predictions on an independent hold-out set. Models were further trained only with data available at PS combined with all biomarkers from tissue microarrays but not TNM (PS dataset). Feature selection was applied with permutation feature importance (PFI) to create reduced datasets with only important variables for predictions. RESULTS AI methods predicted the five-year survival status with a comparable accuracy when trained with the whole dataset (Accuracy: 0.77/0.76/0.76/0.74/0.69 RF/XG/ANN/TN/LR, respectively). When trained without the biomarkers but with complete patient characteristics, including TNM, model predictions did not deteriorate. LR showed the least accurate prediction in any conducted computational experiment. In contrast, models trained only with collected data until PS with biomarkers showed better predictive power compared to excluded biomarkers (whole PS dataset vs. PS dataset without biomarkers; Accuracy: 0.77 vs. 0.70/ 0.79 vs. 0.73/0.75 vs. 0.71/0.72 vs.0.69/0.66 vs. 0.63 RF/XG/ANN/TN/LR). Model predictions with selected features via PFI showed similar results compared to when trained with all features. Important feature overlap of AI methods, when trained with all features, was: pN status, pT status, p16 deletion, and Her2/neu amplification. Feature overlap when trained with the PS dataset was: patient age at the time of surgery, TP-53 mutation, Mesothelin expression, TYMP expression, NANOG expression, IDO expressed on tumor-infiltrating lymphocytes, tumor-infiltrating mast- and NK-cells. CONCLUSION AI can predict the long-term survival of patients with EC. Survival status can be predicted at the time of PS if additional information on the tumor tissue is available. This suggests that individual predictions are possible early in cancer treatment with biomarkers and do not rely on the pathological TNM status after surgery.