GPC-net: a generative-augmented predictive-corrective neural network for small-sample FTIR spectral quantification...

Quantitative analysis in biomedical Fourier-transform infrared (FTIR) spectroscopy faces a fundamental challenge known as the small sample, high dimensionality paradox. This study introduces GPC-Net, a three-stage neural framework developed for robust spectral quantification with limited data. The framework integrates a conditional variational Transformer autoencoder for generating physically realistic synthetic spectra, a Predictive Neural Network (PNN) pretrained on this augmented data to establish a global spectral-to-target mapping, and a Corrective Neural Network (CNN) trained on real calibration samples to perform sample-specific residual correction. We validated GPC-Net for estimating subdural hematoma (SDH) injury time using a rat model and human clinical samples. The model demonstrated superior predictive accuracy over established baselines such as Partial Least Squares (PLS) regression, Artificial Neural Networks (ANN), Random Forest (RF),XGBoost and One-dimensional convolutional neural network (1D-CNN). Evaluations confirmed the high fidelity of the synthetic spectra. Ablation studies established the necessity of each architectural component. SHAP interpretability analysis showed that model decisions are associated with biochemically relevant spectral regions, including the amide I/II bands, and their temporal contribution patterns align with known hematoma aging pathology. The framework also maintained stable performance under label noise. GPC-Net provides an accurate, interpretable, and consistent across the datasets evaluated methodology for small-sample FTIR quantification in biomedical applications.