Robust Machine Learning for Colorectal Cancer Risk Prediction and Stratification

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Colorectal cancer (CRC) is a leading cause of cancer-related mortality in the US.
• Current CRC screening methods lack effectiveness for the general population.
• There is a need for accessible and accurate CRC risk assessment tools.

Purpose of the Study:

• To identify an effective mass screening method for colorectal cancer (CRC) risk.
• To evaluate and compare seven supervised machine learning algorithms for CRC risk prediction.
• To determine the optimal algorithm and data imputation method for CRC risk stratification.

Main Methods:

• Trained and cross-tested seven supervised machine learning algorithms: LDA, SVM, Naive Bayes, Decision Tree, Random Forest, Logistic Regression, and ANN.
• Utilized the National Health Interview Survey (NHIS) and Prostate, Lung, Colorectal, Ovarian Cancer Screening (PLCO) datasets.
• Applied six imputation methods (mean, Gaussian, Lorentzian, one-hot encoding, EM, listwise deletion) for missing data handling.

Main Results:

• The Artificial Neural Network (ANN) with Expectation-Maximization (EM) imputation achieved the highest performance (concordance: 0.70 ± 0.02, sensitivity: 0.63 ± 0.06, specificity: 0.82 ± 0.04).
• The best model demonstrated low misclassification rates: 2% of negative cases as high risk and 6% of positive cases as low risk.
• Kaplan-Meier analysis showed statistically significant separation between low-, medium-, and high CRC-risk groups.

Conclusions:

• An Artificial Neural Network model, optimized with EM imputation, is a highly effective tool for mass colorectal cancer (CRC) risk screening.
• This machine learning approach facilitates early intervention and prevention strategies for CRC in large populations.
• The developed model shows significant potential for improving public health outcomes related to colorectal cancer.