Predicting Duodenal Cancer Risk in Patients with Familial Adenomatous Polyposis Using Machine Learning Model

The aim of this study was to both classify data of familial adenomatous polyposis patients with and without duode- nal cancer and to identify important genes that may be related to duodenal cancer by XGboost model.

The current study was performed using expression profile data from a series of duodenal samples from familial adenomatous polyposis patients to explore variations in the familial adenomatous polyposis duodenal adenoma-carcinoma sequence. The expression profiles obtained from cancerous, adenomatous, and normal tissues of 12 familial adenomatous polyposis patients with duodenal cancer and the tissues of 12 familial adenomatous polyposis patients without duodenal cancer were compared. The ElasticNet approach was utilized for the feature selection. Using 5-fold cross-validation, one of the machine learning approaches, XGboost, was utilized to classify duodenal cancer. Accuracy, balanced accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score performance metrics were assessed for model performance.

According to the variable importance obtained from the modeling, ADH1C, DEFA5, CPS1, SPP1, DMBT1, VCAN-AS1, APOB genes (cancer vs. adenoma); LOC399753, APOA4, MIR548X, and ADH1C genes (adenoma vs. adenoma); SNORD123, CEACAM6, SNORD78, ANXA10, SPINK1, and CPS1 (normal vs. adenoma) genes can be used as predictive biomarkers.

The proposed model used in this study shows that the aforementioned genes can forecast the risk of duodenal cancer in patients with familial adenomatous polyposis. More comprehensive analyses should be performed in the future to assess the reliability of the genes determined.