# Comprehensive machine learning analysis of a radiomics-based model for predicting microsatellite instability in right Colon Cancer

## TL;DR

This study develops a machine learning model using radiomics and clinical data to predict microsatellite instability in right colon cancer, offering a noninvasive preoperative tool for clinical decisions.

## Contribution

The novel contribution is a radiomics-based machine learning model integrated with clinicopathological features for predicting dMMR/MSI-H status in right colon cancer.

## Key findings

- The random forest-based joint model achieved AUC values of 0.99 (training) and 0.97 (test) for predicting dMMR status.
- The model showed good generalizability with an AUC of 0.81 in external validation using left colon cancer data.
- The joint model outperformed radiomics-only and clinical-only models in predictive performance.

## Abstract

The objective of this study was to develop and validate a noninvasive radiomics-based machine learning (ML) model integrated with clinicopathological features for the prediction of microsatellite instability [deficient mismatch repair (dMMR)/microsatellite instability—high (MSI-H)] status in right colon cancer, aiming to provide a preoperative decision-making tool for clinical practice.

A total of 247 patients with right colon cancer [43 dMMR and 204 proficient mismatch repair (pMMR)] who underwent radical resection between January 1, 2017, and 31 December 2024, were enrolled and randomly divided into a training set (70%) and a test set (30%). Preoperative contrast-enhanced computed tomography (CT) images were processed using 3D Slicer for region of interest (ROI) delineation and radiomics feature extraction. The intraclass correlation coefficient (ICC) was used to assess interobserver consistency, while the least absolute shrinkage and selection operator (LASSO) regression method was applied for feature selection. Logistic regression (LR), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost) were used to construct radiomics models. The RF algorithm was selected to build a joint clinicopathological–radiomics model, and patients with left colon cancer served as the external validation set. Receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA) were used to evaluate diagnostic efficiency.

A total of 107 radiomics features were extracted, with 17 stable features retained after ICC filtering (ICC ≥ 0.75) and LASSO regression with 50% cross-validation. The RF algorithm outperformed other models in the radiomics model, with area under the curve (AUC) values of 0.98 in the training set and 0.96 in the test set. The joint model integrating the RF algorithm and clinicopathological variables (e.g., sex, age, tumor long diameter, histological type, pN, pM, pTNM, and differentiation degree) achieved the highest predictive performance, with AUC values of 0.99 (training set) and 0.97 (test set), which were significantly higher than those of the radiomics model and the clinical model alone. External validation with left colon cancer data also showed an AUC of 0.81, indicating good generalizability. The calibration curves demonstrated satisfactory probability prediction, and the DCA confirmed that the joint model provided greater clinical net benefit across the entire threshold probability range.

The RF-based joint clinicopathological–radiomics model exhibited excellent performance in predicting the dMMR status in right colon cancer, with good generalizability across the entire colon. This noninvasive model can serve as a reliable clinical decision support tool to optimize risk stratification and guide early intervention for patients with right colon cancer.