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

Area of Science:

• Machine learning applications in oncology
• Epidemiological modeling for disease risk
• Biostatistics and data science in healthcare

Background:

• Gastrointestinal (GI) cancers pose a significant health challenge, particularly in South Korea.
• Machine learning (ML) models offer potential for early screening and risk identification.
• Class imbalance in prospective cohorts often hinders ML model sensitivity for rare diseases like GI cancer.

Purpose of the Study:

• To evaluate class imbalance mitigation strategies for ML-based GI cancer risk prediction.
• To develop models using noninvasive and minimally invasive predictors.
• To link risk factors to modifiable behavioral and metabolic elements.

Main Methods:

• Analysis of a prospective cohort (n=7652) with 156 GI cancer cases over 14 years.
• Development and comparison of a patient-centered undersampling technique (PCUSTe) against SMOTE, ADASYN, and ENN.
• Implementation of six classifiers with probability correction and evaluation using sensitivity, specificity, AUC, and MCC.

Main Results:

• PCUSTe-trained models showed improved sensitivity, especially with complex classifiers.
• An incrementally trained stochastic gradient descent model achieved high performance (Sensitivity: 0.77, AUC: 0.77).
• PCUSTe enhanced sensitivity in complex models, sometimes at the expense of specificity.

Conclusions:

• Integrating epidemiological principles like covariate frequency matching improved minority class detection.
• Model performance varied by algorithm; threshold adjustment alone sometimes sufficed.
• Selected imbalance mitigation strategies can yield models suitable for early GI cancer risk identification and personalized strategies.