Development and Interpretation of a Clinicopathological-Based Model for the Identification of Microsatellite Instability in Colorectal Cancer. (18th February 2023)

• Record Type: Journal Article
• Authors: Jiang, Zhenxing; Yan, Lizhao; Deng, Shenghe; Gu, Junnan; Qin, Le; Mao, Fuwei; Xue, Yifan; Cai, Wentai; Nie, Xiu; Liu, Hongli; Shang, Fumei; Tao, Kaixiong; Wang, Jiliang; Wu, Ke; Cao, Yinghao; Cai, Kailin
• Other Names: Niu Xing Academic Editor.
• Is Part Of: Disease markers. Volume 2023(2023)
• Publication Date: 2023-02-18
• Subjects: Diagnosis -- Periodicals; Biochemical markers -- Periodicals; Pathology -- Periodicals; 616
• DOI: 10.1155/2023/5178750
• Languages: English
• ISSNs: 0278-0240

Abstract

Abstract : Chemotherapy is not recommended for patients with deficient mismatch repair (dMMR) in colorectal cancer (CRC); therefore, assessing the status of MMR is crucial for the selection of subsequent treatment. This study is aimed at building predictive models to accurately and rapidly identify dMMR. A retrospective analysis was performed at Wuhan Union Hospital between May 2017 and December 2019 based on the clinicopathological data of patients with CRC. The variables were subjected to collinearity, least absolute shrinkage and selection operator (LASSO) regression, and random forest (RF) feature screening analyses. Four sets of machine learning models (extreme gradient boosting (XGBoost), support vector machine (SVM), naive Bayes (NB), and RF) and a conventional logistic regression (LR) model were built for model training and testing. Receiver operating characteristic (ROC) curves were plotted to evaluate the predictive performance of the developed models. In total, 2279 patients were included in the study and were randomly divided into either the training or test group. Twelve clinicopathological features were incorporated into the development of the predictive models. The area under curve (AUC) values of the five predictive models were 0.8055 for XGBoost, 0.8174 for SVM, 0.7424 for NB, 8584 for RF, and 0.7835 for LR (Delong test, P value < 0.05). The results showed that the RF model exhibited the best recognition ability and outperformed the conventional LR method in identifying dMMR and proficient MMR (pMMR). Our predictive models based on routine clinicopathological data can significantly improve the diagnostic performance of dMMR and pMMR. The four machine learning models outperformed the conventional LR model.