Immune-Enhanced Machine Learning Approach for Early Detection of Precancerous Colorectal Neoplasia: Insights from Biomarkers in Routine Health Checkups
Menée à partir de données portant sur 6 146 adultes asymptomatiques ayant bénéficié d'une coloscopie et de tests de biomarqueurs immunitaires lors d'évaluations de santé de routine, cette étude examine l'intérêt de modèles d'apprentissage automatique intégrant des données de marqueurs immunitaires pour détecter précocement des néoplasies colorectales précancéreuses
Background and Aims : Current screening strategies for colorectal cancer (CRC) rely on colonoscopy, an invasive procedure with limited capacity to address individual risk. There is growing interest in integrating noninvasive immune biomarkers to improve early detection of precancerous colorectal neoplasia (CRN) in asymptomatic individuals. This study aimed to evaluate the predictive utility of immune-related markers, particularly natural killer cell activity (NKA), in enhancing risk stratification through machine learning approaches.
Methods : Data from 6,146 asymptomatic adults who underwent colonoscopy and immune biomarker testing during routine health evaluations were retrospectively analyzed. A range of machine learning models were developed and evaluated via stratified cross-validation. Additional analyses included longitudinal follow-up of false-positive cases and ablation study to evaluate the importance of immune-related variables. Sex-stratified analyses were conducted.
Results : Integration of immune biomarkers improved prediction of precancerous CRN, with TabNet achieving an area under the curve (AUC) of 0.734 for advanced neoplasms and 0.739 for high risk polyps. Models demonstrated high negative predictive value, supporting their utility for screening triage. Individuals misclassified as having advanced neoplasm despite normal colonoscopy findings exhibited a significantly elevated prevalence of neoplasia on follow-up. Removal of immune variables led to substantial declines in model performance. Notably, the predictive value of NKA was stronger in females, indicating sex-specific immunological dynamics.
Conclusion : Immune-informed machine learning models provide a noninvasive and individualized strategy for CRC screening. The findings support the clinical value of incorporating immune markers, particularly NKA, into risk stratification frameworks and highlight the need for sex-specific approaches in surveillance planning.
European Journal of Cancer , article en libre accès, 2025