Using Synthetic Gaussian Noise to Explore Stochastic Resonance in Cancer Subtype Classification

Stochastic resonance (SR) allows models to detect weak signals or features that might otherwise be lost in noise, thereby improving the classification performance in Artificial Intelligence (AI) models. This study represents one of the first applications of SR in enhancing AI-driven cancer diagnostics. We compared eight common AI classification models (random forest, logistic regression, Naive Bayes, XGBoost, support vector machine, gradient boost, AdaBoost, and K-Nearest Neighbor) using performance metrics to evaluate their effectiveness in distinguishing cancer subtypes at varying noise levels. We trained our machine learning models using a dataset published by the University of California, Irvine’s Machine Learning Repository, which consists of 20,531 Ribonucleic Acid sequence gene expression features and 801 samples representing the five cancer subtypes. While previous studies on this dataset focused on feature selection, our study explores the controlled addition of synthetic Gaussian noise at five different levels through SR to uncover weak signals in the data. Furthermore, we selected 80 significant genes by evaluating the Analysis of Variance f-values and removing constant features. Our results indicate that models enhanced with SR can classify gene expression with high accuracy (93-98%) when noise levels are increased to $\sigma =4$ . Logistic regression and support vector machine models performed the best in noise-free conditions. However, model accuracy decreased when noise levels exceeded $\sigma =4$ , declining to 83-93% when $\sigma =8$ across all models, showing machine learning vulnerability to unrestrained noise. The results of this study suggest that controlled noise implementation can enhance cancer diagnostic accuracy through SR.