Neural Network-Based Composite Risk Scoring for Stratification of Fecal Immunochemical Test-Positive Patients in...

Background: Colorectal cancer (CRC) remains a major cause of cancer-related mortality worldwide, underscoring the need for more efficient and resource-conscious screening strategies. Methods: We screened 51,437 individuals (50-74 y) in South-West Oltenia, Romania, with FIT values of ≥20 µg Hb/g. Of the 2825 FIT-positive individuals, 1550 completed colonoscopy, and we recorded their age, sex, residence, education, comorbidities, medications, and FIT values. After imputing < 8% missing data via multiple imputation, we reduced dimensionality with an autoencoder (ReLU, dropout 0.5, L2, 100 epochs, batch 32) and applied K-Means clustering (k = 5). The following are examples of actionable clusters: Cluster 0 ("High-FIT malignant"): FIT > 200 µg/g, age > 65, diabetes; Cluster 2 ("Low-risk mixed"): FIT 100-199 µg/g, age < 60, no comorbidities; Cluster 3 ("Intermediate-risk older"): FIT 150-200 µg/g, ≥3 comorbidities, rural. Cluster labels were then predicted by a feed-forward neural network (64-32 neurons, dropout 0.6) and validated via 5-fold cross-validation plus a temporal hold-out. Results: Five distinct patient clusters were identified, enabling the development of a composite risk score. Notably, Cluster 0, characterized by elevated FIT levels, exhibited a malignancy rate of 50.91%, while the overall CRC diagnostic rate among colonoscoped patients was approximately 13.87%. This stratification model enhances the diagnostic yield by prioritizing high-risk patients for urgent colonoscopy and sparing low-risk individuals from unnecessary invasive procedures. Conclusions: The AI-driven composite risk score offers a refined framework for CRC risk stratification and optimized resource allocation. Its implementation can lead to earlier detection of advanced lesions, thereby improving patient outcomes. Further external validation on independent cohorts and regions is essential to confirm its broad utility, with potential future integration of additional biomarkers (e.g., genetic or omics-based indicators) to further enhance predictive accuracy.