Data Augmentation and Machine Learning Classification of Serum SERS Spectra in Patients with Non‑Infectious Diseases

Abstract

In this study, we evaluated spectral augmentation strategies to mitigate class imbalance and limited sample sizes and improve the robustness of partial least squares discriminant analysis (PLS-DA) classifiers, using ROC AUC as the primary performance metric. Among simple transformations, spectral shifting by ±3 cm⁻¹ with cumulative addition to the original dataset yielded optimal results: ROC AUC remained high (0.983 ± 0.003) while the standard deviation significantly decreased compared to the baseline (0.982 ± 0.022) on a sixfold augmented dataset. In contrast, MixUp (α = 0.1) and CutMix (cut_ratio = 0.3) at 10× augmentation degraded performance (ROC AUC 0.631–0.663) due to structural distortion of spectral features. Generative approaches demonstrated superior efficacy. A variational autoencoder (VAE) with a four-layer convolutional encoder (32–256 feature maps), residual blocks, and a 32-dimensional latent space, coupled with a decoder comprising three fully connected layers followed by upsampling, substantially improved classification accuracy and reduced metric variance. Similarly, a Wasserstein GAN (WGAN) with a two-layer generator (512/1024 units, tanh output) and a three-layer critic, optimized using RMSprop with a 5:1 update ratio, stabilized ROC AUC at 0.97. Conversely, SVD-based augmentation (retaining 95% variance) proved ineffective, reducing ROC AUC to 0.565 ± 0.012 at 10× expansion. These findings demonstrate that spectral shifting (±3 cm⁻¹), VAE, and WGAN preserve spectral integrity and enhance model stability, offering practical solutions for robust SERS-based serum diagnostics in imbalanced clinical datasets. This research was funded by Russian Science Foundation, grant number 25-75-00146; https://rscf.ru/project/25-75-00146/.

Speaker

Khristoforova Yulia
Samara National Research University
Russian Federation

Discussion

Ask question