Colorectal Cancer Detected by Machine Learning Models Using Conventional Laboratory Test Data

Abstract

Background: Current diagnostic methods for colorectal cancer (CRC) are colonoscopy and sigmoidoscopy, which are invasive and complex procedures with possible complications. This study aimed to determine models for CRC identification that involve minimally invasive, affordable, portable, and accurate screening variables. Methods: This was a retrospective study that used data from electronic medical records of patients with CRC and healthy individuals between July 2017 and June 2018. Laboratory data, including liver enzymes, lipid profiles, complete blood counts, and tumor biomarkers, were extracted from the electronic medical records. Five machine learning models (logistic regression, random forest, k-nearest neighbors, support vector machine [SVM], and naïve Bayes) were used to identify CRC. The performances were evaluated using the areas under the curve (AUCs), sensitivity, specificity, positive predictive values (PPV), and negative predictive values (NPV). Results: A total of 1164 electronic medical records (CRC patients: 582; healthy controls: 582) were included. The logistic regression model achieved the highest performance in identifying CRC (AUC: 0.865, sensitivity: 89.5%, specificity: 83.5%, PPV: 84.4%, NPV: 88.9%). The first four weighted features in the model were carcinoembryonic antigen (CEA), hemoglobin (HGB), lipoprotein (a) (Lp(a)), and high-density lipoprotein (HDL). A diagnostic model for CRC was established based on the four indicators, with an AUC of 0.849 (0.840-0.860) for identifying all CRC patients, and it performed best in discriminating patients with late colon cancer from healthy individuals with an AUC of 0.905 (0.889-0.929). Conclusions: The logistic regression model based on CEA, HGB, Lp(a), and HDL might be a powerful, noninvasive, and cost-effective method to identify CRC.

Keywords: clinical laboratory techniques; colorectal cancer; diagnosis; logistic regression; machine learning.

Figure 1.

Flowchart of the colorectal cancer (CRC) identification model.

Figure 2.

Matrix of the Spearman correlation coefficients. All pairs of variables included in the models were tested using the Spearman correlation. For the variable pairs in which correlation coefficients >0.5, the one with the less weight coefficient in the principal component analysis (PCA) was deleted from feature collection.

Figure 3.

The weight coefficients of the logistic regression model (the model with the highest accuracy) for colorectal cancer (CRC) diagnosis. The first four weighted features in the logistic regression model were carcinoembryonic antigen (CEA), hemoglobin (HGB), lipoprotein (a) (Lp(a)), and high-density lipoprotein (HDL).

Figure 4.

Receiver operating characteristic (ROC) curve for colorectal cancer (CRC) diagnosis using logistic regression models: CEA alone, CEA + hemoglobin (HGB)  + Lp(a), CEA + HGB + Lp(a) + HDL, and CEA + HGB + Lp(a) + HDL + ALT.