A diagnostic model based on cell-free DNA (cfDNA) methylation markers demonstrated high accuracy in detecting both colorectal cancer and precancerous lesions.  

“The performance positions [the model] as a promising tool for reliable, early screening in clinical practice,” said lead author Li-Yue Sun, MD, PhD, of the Department of Health Management Centre, The First Affiliated Hospital of Jinan University in Guangzhou, China, during a presentation at the 2026 ASCO Breakthrough Meeting (Abstract 94). 

“Incorporation of low-cost cfDNA assays significantly enhances the model’s ability. This key feature unlocks its potential for widespread clinical application, making advanced cancer screening accessible even in resource-limited health-care settings,” Dr. Sun added.  

Background 

“Colorectal cancer is the leading cause of cancer deaths worldwide. However, it is a disease [for which screening can] change outcomes,” Dr. Sun said.  

Current colorectal cancer screening methods include fecal immunochemical testing, which has a sensitivity of 85%, specificity of 95%, and patient acceptability of approximately 90%; multitarget stool DNA testing, with a sensitivity of 92%, specificity of 90%, and patient acceptability of 85%; and colonoscopy, with a sensitivity of 98%, specificity greater than 99%, and patient acceptability of 75%.  

The researchers sought to develop a more acceptable and accessible screening approach for colorectal cancer.  

Study Methods 

They integrated cfDNA testing with multiple detection methods—including electrochemical detection, colloidal gold detection, colloidal gold strip testing, and traditional biomarker testing—and applied machine learning to develop a diagnostic model for detecting intestinal adenomas and colorectal cancer.  

The study enrolled 1,373 individuals across four groups: patients with colorectal cancer (n = 261), those with precancerous lesions (n = 312), high-risk individuals (n = 400), and 400 age- and sex-matched healthy controls. Dr. Sun noted that all clinical data underwent rigorous standardization.  

Model Methods 

The researchers completed a comprehensive assessment of 15 machine learning models, including traditional logistic regression, support vector machine, random forest, XGBoost, and AdaBoost approaches.  

Model performance was assessed using area under the curve, sensitivity, and specificity, with model selection based on both predictive performance and clinical relevance.  

The investigators also performed a feature importance analysis using the XGBoost model to evaluate the contribution of individual predictive variables.  

Finally, they incorporated electrochemical cfDNA adsorption and colloidal gold–based cfDNA adsorption detection methods into the selected model.  

Model Performance 

The AdaBoost model was selected as the optimal approach. For colorectal cancer detection, it achieved an area under the curve of 0.999, with a sensitivity of 98.1% and a specificity of 99.8%. For the detection of precancerous lesions, the model achieved an area under the curve of 0.843, with a sensitivity of 89.1% and a specificity of 62.6%, a performance the investigators considered clinically meaningful. 

“This result indicates that the model is highly accurate for detecting cancer, [though less effective] in precancerous lesions, which remains a challenging clinical target,” Dr. Sun said.  

The model’s top predictive features included the CA19-9 biomarker, smoking index, white blood cell count, glucose, cholesterol, triglycerides, carcinoembryonic antigen, the CA72-4 biomarker, age, and neutrophil-to-lymphocyte ratio.  

Dr. Sun reiterated that the model addresses several limitations of traditional screening approaches by offering a noninvasive, accurate method that integrates molecular and clinical data. 

Going forward, the researchers plan to prospectively validate the model and evaluate its integration into routine clinical workflows.  

DISCLOSURES: The research was funded by the National Natural Science Foundation of China. The study authors reported no financial relationships.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.