Artificial Intelligence-based Colon Cancer Prediction by Identifying Genomic Biomarkers

Abstract

Colon cancer is the third most common type of cancer worldwide. Because of the poor prognosis and unclear preoperative staging, genetic biomarkers have become more important in the diagnosis and treatment of the disease. In this study, we aimed to determine the biomarker candidate genes for colon cancer and to develop a model that can predict colon cancer based on these genes. Material and Methods: In the study, a dataset containing the expression levels of 2000 genes from 62 different samples (22 healthy and 40 tumor tissues) obtained by the Princeton University Gene Expression Project and shared in the figshare database was used. Data were summarized as mean ± standard deviation. Independent Samples T-Test was used for statistical analysis. The SMOTE method was applied before the feature selection to eliminate the class imbalance problem in the dataset. The 13 most important genes that may be associated with colon cancer were selected with the LASSO feature selection method. Random Forest (RF), Decision Tree (DT), and Gaussian Naive Bayes methods were used in the modeling phase. Results: All 13 genes selected by LASSO had a statistically significant difference between normal and tumor samples. In the model created with RF, all the accuracy, specificity, f1-score, sensitivity, negative and positive predictive values were calculated as 1. The RF method offered the highest performance when compared to DT and Gaussian Naive Bayes. Conclusion:In the study, we identified the genomic biomarkers of colon cancer and classified the disease with a high-performance model. According to our results, it can be recommended to use the LASSO+RF approach when modeling high-dimensional microarray data.

A Fecal-Microbial-Extracellular-Vesicles-Based Metabolomics Machine Learning Framework and Biomarker Discovery for Predicting Colorectal Cancer Patients

“…Therefore, we considered only metabolites overlapping the two different statistical approaches for further analysis (FDR-corrected p -value < 0.05 and AUC > 0.70). Multivariate analyses were performed using the ROC curve method with biomarker candidate metabolites based on linear support vector machine (SVM) [ 17 ], partial least squares discrimination analysis (PLS-DA) [ 18 ], and random forest (RF) [ 19 ] algorithms. These methods have proved to be robust for high-dimensional data and are widely used for other types of ‘omics’ data analysis.…”

Exploring obesity, physical activity, and digital game addiction levels among adolescents: A study on machine learning-based prediction of digital game addiction

“…Synthetic Minority Over-sampling Technique for Nominal and Continuous (SMOTE-NC) Gök and Olgun (2021) was used to eliminate the class imbalance problem. Class imbalance problem, when working with real-life data, this problem is highly prevalent and can be defined as a state of imbalance when there are significantly more cases belonging to the majority class than those belonging to the minority class (Paksoy and Yağin, 2022). Because ML techniques like logistic regression might be biased toward the majority class, which causes issues with under or overfitting, balanced data is crucial.…”

Genomic Biomarkers of Metastasis in Breast Cancer Patients: A Machine Learning Approach

“…An algorithm based on decision-tree (DT) and gradient-boosting (GB), XGBoost is a faster running algorithm compared to GB algorithms, with different regularization penalties to avoid overfitting [18]. NB is an algorithm based on conditional probability, which is assumed to be equal and independent from each other in the classification of all attributes based on conditional probability [19].…”