AI-based complete blood count model for colorectal cancer detection.

e15693 Background: Early colorectal cancer (CRC) detection is crucial for effective treatment; however, traditional screening methods face challenges. Colonoscopy, though effective, has limited availability, especially in resource-constrained settings. Conversely, fecal occult blood tests are cost-effective and widely available but suffer from low adherence. These barriers underscore the need for innovative approaches to support CRC screening efforts. Our retrospective study proposes developing a machine learning model based on complete blood count (CBC) tests as a risk stratification tool for CRC. By identifying high-risk individuals, this method could facilitate resource prioritization and active case-finding in at-risk populations, complementing existing screening programs and improving overall accessibility. Methods: We analyzed CBC tests from 7,588 individuals (3,990 females, 52,58%; 3,598 males, 47,42%) aged 45-75 who underwent colonoscopy or biopsies within six months of their CBC test. Among them, 468 (6,17%) were identified as cases confirmed by anatomopathological tests. The remaining 7,120 (93,83%) individuals were classified as controls, with typical colonoscopy results and without polyps, neoplasms, or other abnormalities. The database was divided into training (80%) and validation (20%) sets. The model was developed using ridge regression. Results: Descriptive analysis of all CBC biomarkers, CBC-derived ratios, and age revealed significant differences between controls and CRC cases across all features ( P < 0.001), except for lymphocytes. Using a feature selection methodology, the markers RDW, leukocytes, hemoglobin, and age were incorporated in a ridge regression model, achieving an AUC of 0.75 (95% CI: 0.74–0.75). The model was trained on a combined dataset to assess overall performance without introducing sex-based biases. It was then tested separately on male and female subsets, yielding an AUC of 0.76 (95% CI: 0.75–0.78) for males and 0.75 (95% CI: 0.74–0.75) for females, indicating consistent performance across genders. Explainability analysis revealed that increased age, higher RDW, elevated leukocyte counts, and lower hemoglobin levels were linked to a higher risk of CRC. Due to the imbalance in our dataset, we selected the threshold that maximized the balanced accuracy, resulting in a sensitivity of 0.72, specificity of 0.66, accuracy of 0.66, balanced accuracy of 0.69, negative predictive value (NPV) of 0.97, and positive predictive value (PPV) of 0.13. Conclusions: Our findings suggest that our model can contribute to the early detection of CRC, potentially serving as a valuable risk stratification tool for screening. However, the absence of external validation is a limitation of this study, requiring further research to confirm its generalizability and clinical impact.