Machine Learning-Based Prediction Model for Early-Onset Colorectal Cancer Risk: A Systematic Review and Meta-Analysis

Through systematic review and meta-analysis, we explored the performance of machine learning (ML) risk prediction models for early-onset colorectal cancer (EOCRC) to enhance model development and application. Following preregistration of the study protocol in PROSPERO (CRD42024606785), we searched PubMed, Embase, Web of Science, Cochrane Library, WanFang Data, VIP, SinoMed, and the CNKI databases for studies on constructing predictive EOCRC based on ML methods. The prediction model Risk of Bias Assessment Tool + Artificial Intelligence (PROBAST + AI) tool was used to evaluate the quality, bias risk, and applicability of the included studies. The Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis + Artificial Intelligence (TRIPOD + AI) reporting standard was applied to assess the reporting quality of these studies. We conducted subgroup analysis and sensitivity analysis. This study followed the PRISMA reporting guidelines. A total of 2809 studies were retrieved, and 7 studies were included. The training set size ranged from 987 to 21,700 patients, with an average of 5037 per model. XGBoost was the most frequently used modelling method. Overall TRIPOD + AI compliance was 70.5%. In model development, 5 studies were rated as high-quality concerns (2 unclear), and 5 studies were rated as high adaptability concerns (2 unclear). In model evaluation, 6 studies were rated as high-risk bias (1 unclear), and 5 studies were of concern for high adaptability (2 unclear). The top 5 most common predictors include age, gender, BMI, diabetes and smoking. Meta-analysis showed a pooled AUC of 0.84 (95% CI, 0.77-0.90; I² = 99.3%, P < .001). Pooled sensitivity and specificity were 0.75 (95% CI, 0.59-0.86; I² = 94.8%) and 0.87 (95% CI, 0.71-0.94; I² = 95.6%), respectively (both P < .001 for heterogeneity). This study demonstrates that ML models show outstanding discriminatory capabilities in predicting the risk of EOCRC, supporting their potential application in identifying high-risk individuals for targeted prevention and monitoring. In the future, prospective design and standardized reporting are needed to enhance reliability and clinical translation.