Abstract
This study focuses on the classification of optical fiber biosensor signals measured at seven different cancer cell concentrations. Data scarcity in biological signal acquisition was addressed by generating synthetic data using a conditional variational autoencoder (CVAE). Five classification algorithms were trained and evaluated to distinguish between the seven concentrations. The Random Forest model achieved the highest performance, with 77% accuracy on the original dataset, demonstrating the ability to capture concentration-dependent signal variations. However, a comparison of different proportions of real and synthetic data shows that increasing the proportion of synthetic data decreases accuracy. These results indicate that while generative models offer a potential solution to data scarcity, their effectiveness depends on their ability to accurately capture the underlying structure of biosensor signals.