26 PREDICTION OF PREOPERATIVE CHEMOTHERAPY FOR ESOPHAGEAL CANCER USING ARTIFICIAL INTELLIGENCE

Currently, preoperative chemotherapy is usually done for advanced esophageal cancer according to guidelines. It is still difficult to predict the effect of chemotherapy before the chemotherapy. If the effect would be poor, the patient may miss the possibility of treatment. We have been examining the possibility of predicting the chemotherapy effect based on test results before and during chemotherapy. In this study we used artificial intelligence to predict the effect of chemotherapy. Methods 80 patients who underwent preoperative chemotherapy for esophageal cancer in our department from 2012 to 2019. We predict the chemotherapy effect according to age, sex, type of chemotherapy (FP/DCF), number of chemotherapy, T, M, N, blood examination results before chemotherapy as WBC, neutrophil count, lymphocyte count, hemoglobin, platelets, total protein, Albumin, CRP, IgG, IgA, IgM, IL6, creatinine, lowest WBC, neutrophil and lymphocyte count, during chemotherapy. The effect of chemotherapy was classified: CR, PR, and PD. Prediction were performed using MatLab R2019b classification learner and Neural Net Pattern Recognition. Results The actual effect was CR/PR/SD/PD = 0/33/42/5 cases. The correct answer rate was 66.3% for the optimized tree model, 66.3% for the optimized Naïve Bayesian model, 67.5% for the optimized SVM model, and 56.3% for the optimized ensemble model. With optimized SVM, the sensitivity to predict PD was 20%, specificity was 94.5%, and AUC was 0.83. Moreover, the sensitivity to predict PR was 75.8%, and the specificity was 76.6%. With Neural Net Pattern Recognition, the correct answer rate was 90%, and the sensitivity to predict PD was 20%, specificity was 94.7%. Conclusion In this study, Neural Net Pattern Recognition as a deep learning model could predict the effect of preoperative chemotherapy more accurately than that of classification learner. The poor prediction of PD was due to the small number of teacher models, and the prediction of PR with many teacher models was more accurate. We will add more teacher models, and establish a prediction model that can be used clinically for both sensitivity and specificity.