Implementation and Evaluation of Models for Synthetic Generation of Spectral Signals to Balance Metabolomic Data

Although the advance of Artificial Intelligence (AI) in medicine is significant, its potential impact depends on access to quality data. A particular challenge is class imbalance in medical data sets. This imbalance can introduce biases into AI models that reduce their ability to accurately identify and/or diagnose less common health conditions. To address this problem, synthetic data generation techniques have emerged as a promising strategy. In this project, it is proposed to balance a dataset with patients’ metabolic profiles with and without conditions associated with the Metabolic Syndrome (MetS), through the synthetic data generation of nuclear magnetic resonance (NMR) spectral signals. Two deep learning models (GAN and VAE) have been implemented for the generation of synthetic NMR spectral signals. The synthetic signals have been subjected to analysis using various metrics to evaluate their similarity to real signals from blood samples. The used metrics to compare the shape and trend of the signals and to examine correlations between metabolites include the Pearson correlation coefficient, Dynamic Time Warping (DTW) and heat maps. The best results were achieved with a Variational Autoencoder (VAE) model, for which the correlation results between the real and synthetic signals have been consistently high, and DTW analysis has shown significant similarity between the signals in terms of shape. Overall, it has been shown that synthetic NMR spectral signal generation can be applied for dataset balance, proving that the inclusion of synthetic data from a minority class can improve the analysis of unbalanced datasets, although it is essential to explore strategies to assess further these results.