Near-infrared spectral data augmentation for underrepresented classes using generative adversarial networks to...

This study addresses the challenge of imbalanced datasets in near-infrared spectroscopy classification of urinary stones, specifically the underrepresentation of phosphate stones that leads to suboptimal predictive performance. We developed a data augmentation approach combining generative adversarial networks (GANs) with Soft Independent Modeling of Class Analogies (SIMCA) filtering to generate synthetic phosphate spectra. A total of 400 artificial spectra were generated, showing visual and PCA-based similarity to real data. Support vector machine (SVM) classification models were trained and evaluated on both original and augmented datasets to distinguish between urate, phosphate, and oxalate stones. Different percentages of the real samples from the total dataset were used in each training set (30-70%), and 30% of real samples were allocated for validation only. For the augmented data, precision for phosphates increased by 5-10% depending on the size of the training set. For oxalates and urates, which were not oversamples, precision and recall reached 65-89% and 71-88%, respectively. The inclusion of artificially generated phosphate spectra improves classification performance on an independent test set when distinguishing between urate, phosphate, and oxalate stones. The GAN + SIMCA algorithm can be recommended for similar medical tasks where the imbalance of data is implied, being efficient even for the limited amount of real data.