Machine Learning Classification of Augmented Raman Serum Spectra for Chronic Heart Failure Detection

Abstract

In this work, we propose a promising non-invasive approach based on the Raman spectroscopy of human serum for chronic heart failure (CHF) diagnosis. Due to the limited sample size, which limits the performance of machine learning classifiers, this study explores data augmentation techniques to improve the classification of low- and high-grade CHF spectra using k-nearest neighbor (kNN) and partial least squares discriminant analysis (PLS-DA) algorithms. Raman spectra of 151 patients with CHF of the different stages were acquired at 532 nm excitation from serum samples collected at Samara City Clinical Hospital. Two augmentation approaches were systematically evaluated: (1) combined approach based on the linear spectral transformations (wavenumber shifting ±1–2 cm⁻¹, intensity stretching
0.9-1.1) and (2) Wasserstein Generative Adversarial Network (WGAN)-based synthetic spectrum generation, expanding training datasets 10-fold while preserving physiochemical realism. Augmentation based on linear spectral transformations yielded algorithm-specific results: kNN showed no significant receiver operating characteristic area under curve (ROC AUC) improvement (0.67 ± 0.11 original vs. 0.67–0.69 augmented), while PLS-DA achieved statistically significant gains (0.71 ± 0.11 vs. 0.80–0.81; z = 2.3–2.6, p < 0.05). WGAN augmentation proved superior across both methods, with k-NN reaching 0.74 ± 0.08 and PLS-DA achieving 0.83 ± 0.09. These findings establish WGAN as an optimal augmentation strategy for Raman-based CHF classification, achieving clinically relevant performance (ROC AUC > 0.80) from limited cohorts while enabling biomarker identification for cardiovascular diagnostics.