Advancing Gastrointestinal Cancer Risk Prediction With Patient-Centered Machine Learning: Machine Learning Modeling Study.

Background: Gastrointestinal (GI) cancers are a significant health concern in South Korea. Recently, machine learning (ML) models have emerged as powerful tools to support early screening efforts and identify people at risk before disease onset. However, the low incidence of GI malignancies in prospective cohorts leads to severe class imbalance, often causing ML models to favor the majority "healthy" class at the expense of clinical sensitivity. Objective: This study aimed to evaluate class imbalance mitigation strategies and develop ML-based GI cancer risk prediction models using noninvasive and minimally invasive predictors linked to modifiable behavioral and metabolic risk factors. Methods: We analyzed a prospective cohort (n=7652) with 156 incident GI cancer cases (2%) identified over a 14-year follow-up period. The data were randomly split into training (5356/7652, 70%) and testing (2296/7652, 30%) sets. To address class imbalance while preserving observed population structure, we developed a patient-centered undersampling technique (PCUSTe) based on the logic of frequency-matched case-control studies. PCUSTe was compared with commonly used resampling approaches, including synthetic minority oversampling (SMOTE), adaptive synthetic sampling (ADASYN), and SMOTE with edited nearest neighbors (ENN). Six classifiers were implemented, including both batch and incremental training variants. To account for the prior shift introduced by resampling, probability correction was applied. Model performance was evaluated on the independent test set using a classification threshold equal to the observed event proportion (cumulative incidence) in the training data and then across thresholds reflecting incidence values between 1% and 5%. Primary performance metrics included sensitivity, specificity, Matthews correlation coefficient, and area under the receiver operating characteristic curve (AUC). Results: Models trained using PCUSTe demonstrated improved sensitivity compared with standard resampling techniques, particularly for more complex classifiers. The incrementally trained stochastic gradient descent model achieved the highest overall performance trained on PCUSTe data with a sensitivity of 0.77 (95% CI 0.64-0.89), specificity of 0.65 (95% CI 0.63-0.67), AUC of 0.77 (95% CI 0.70-0.84), and Matthews correlation coefficient of 0.12 (95% CI 0.08-0.16). In contrast, logistic regression achieved balanced performance without resampling (sensitivity 0.70, 95% CI 0.57-0.83; specificity 0.71, 95% CI 0.69-0.72; AUC 0.75, 95% CI 0.68-0.82). Our results showed that PCUSTe primarily enhanced sensitivity in more complex models at the expense of specificity. Conclusions: Integrating epidemiological principles, including covariate frequency matching and threshold selection based on the observed cumulative incidence in the training data, improved minority class detection in GI cancer risk prediction. However, model performance varied by algorithm, and in some cases, decision threshold adjustment alone achieved comparable or superior results to data resampling. These findings highlight the importance of carefully selecting imbalance mitigation strategies based on modeling objectives. The resulting models achieved sensitivity levels that may be suitable for early risk identification in cohort settings and could contribute to personalized risk stratification and targeted prevention or screening strategies.