Developing a machine-learning model to enable treatment selection for neoadjuvant chemotherapy for esophageal cancer.

Although neoadjuvant chemotherapy with docetaxel + cisplatin + 5-fluorouracil (CF) has been the standard treatment for stage II and III esophageal cancers, it is associated with severe adverse events caused by docetaxel. Consequently, this study aimed to construct a prognostic system for CF regimens, especially for locally advanced esophageal cancers. Biopsy specimens from 82 patients treated with the CF regimen plus radical surgery were analyzed. Variants in 56 autophagy- and esophageal cancer-related genes were identified using targeted enrichment sequencing. Overall, 13 single-nucleotide variants, including 8 non-synonymous single-nucleotide variants, were identified as significantly associated with esophageal cancer recurrence (p < 0.05). Particularly, variants of ATG2A p.R478C and ULK2 splice-site also showed significant differences in recurrence-free and overall survival. Subsequently, machine learning was used to construct a model for predicting esophageal cancer recurrence based on 21 features, including eight patient characteristics. A Naive Bayes machine-learning model was shown to be highly reliable for predicting esophageal cancer recurrence with an accuracy of 0.88 and an area under the curve of 0.9. We believe that our results provide useful guidance in the selection of neoadjuvant adjuvant chemotherapy, including avoidance of docetaxel.