Stacked random forest model for colorectal cancer detection using complete blood counts

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Colorectal cancer (CRC) screening adherence is suboptimal in China due to cost and adverse event concerns.
• Complete blood count (CBC) data offers a potential low-cost solution for prioritizing colonoscopy referrals.

Purpose of the Study:

• To develop and validate a machine learning model using CBC data for predicting CRC risk.
• To enhance the efficiency of CRC screening and improve colonoscopy referral decisions.

Main Methods:

• A multicenter study utilized CBC data from participants within three months of colonoscopy.
• A stacking machine learning model was developed using 24 CBC features and 5 combined components (Types A and B CBC data).
• Model performance was assessed using area under the curve (AUC), specificity, and sensitivity.

Main Results:

• The study analyzed 1795 CRC cases and 26,380 cancer-free individuals.
• External validation showed 80.3% specificity and 65.2% sensitivity for the CRC prediction model.
• The model achieved 41% sensitivity for Stage I CRC and 57.6% for Stages I-III combined.

Conclusions:

• CBC testing is a low-cost, accessible tool for preliminary CRC risk assessment.
• The developed machine learning model can aid in colonoscopy referral decisions, improving CRC screening efficiency.