Enhancing Prediction Accuracy in Gastric Cancer Using High-Confidence Machine Learning Models for Class Imbalance

Abstract

Gastric Cancer (GC) diagnosis and prognosis present significant challenges in the clinical industry. To address the issue of low prediction accuracy resulting from imbalanced positive and negative GC cases, this study proposes a medical Decision Support System (DSS) based on supervised Machine Learning (ML) methods. Four ML models, including Naï ve Bayes (NB), Logistic Regression (LR), and Multilayer Perceptron (MLP), were employed in this study. The impact of data imbalance on GC prediction was assessed through two procedures. Among the ML models, the MLP model demonstrated the best performance in weighted GC prediction, achieving a sensitivity of 0.930 and a Positive Predictive Value (PPV) of 0.932 for balanced predictions, and a sensitivity of 0.918 and a PPV of 0.908 for unbalanced predictions. The NB model showed promise in handling the data imbalance issue, achieving a sensitivity of 0.722 and a PPV of 0.420 on the unbalanced dataset. Additionally, a DSS was developed specifically for the NB and LR models to improve prediction accuracy. The proposed method significantly improved the sensitivity of optimistic GC case prediction, with the Naï ve Bayes model achieving a sensitivity of 0.936 and the Logistic Regression model achieving a sensitivity of 0.8306. These improvements enhance the reliability and efficiency of GC diagnostics, offering valuable decision support in healthcare. This research provides insights into addressing class imbalance in GC likelihood prediction and has potential implications for clinical practice.