Scalable Synthetic Data Generation for HDLSS SERRS cTnI Detection via Conditional GMMs

The diagnosis of Myocardial Infarction (MI) can be enhanced through label-free biosensing techniques like Surface-Enhanced Resonance Raman Spectroscopy (SERRS). Machine learning is hindered by data scarcity, excessive dimensionality,<br/>and class imbalance. This paper presents a scalable and interpretable pipeline for generating synthetic data for high-dimensional, low-sample-size (HDLSS) SERRS data. The method combines class-conditional Gaussian Mixture Models (GMMs), physiological renoising through Savitzky–Golay filtering and Gaussian noise, and manifold-constrained filtering using UMAP-KNN to maintain physiological validity. A composite score that integrates fidelity (Wasserstein), diversity (DTW), and generalisability (UMAP acceptance) is used to rank synthetic data configurations during generation. Experimental results indicate that classifiers trained over 5x-scaled synthetic data attain F1 scores of up to 82.35% and ROC AUC of 1.00, surpassing the performance of models trained solely on real data. This lightweight, modular approach offers a scalable alternative to<br/>deep generative models in the context of clinical biosensing.