Augmentation of ATR-FTIR Spectra in Biomedical Fluids: A Baseline-Centered, Geometry-First Evaluation with Stability and Synthetic Data Quality Control

Augmenting infrared spectra is increasingly used to mitigate data scarcity, yet on small clinical datasets it may also create misleading improvements if evaluated without a strict baseline and synthetic quality control (QC). We treat augmentation as a methodological experiment and compare a baseline (training on original data only) against train-only spectrum-level augmentations (Gaussian noise, spectral shift and Mixup) in three biomedical use cases: 1) gingival crevicular fluid (GCF, n=18) with multi-factor clinical annotation, 2) saliva COVID-19 screening (patient-level protocol), and 3) saliva diabetes screening (holdout protocol). Beyond supervised metrics (Recall/F1 for pathology, PR-AUC, specificity), we evaluate probability calibration (Brier score and expected calibration error, ECE) and, crucially, the geometry of the feature space via PCA. For the small-n GCF dataset, we quantify factor–coordinate association strength using ANOVA- and correlation-based measures and demonstrate that augmentation can rotate the PCA basis; therefore, comparisons must rely on best-PC or top-k criteria rather than fixed PC1/PC2. The key finding on the small-n GCF dataset is geometric rather than clinical or biological. Within a predefined narrow spectral window frequently used in prior GCF studies, classic augmentation redistributes variance across higher-order PCA components, leading to an increase in the maximum factor–PC association metric (Δmax R² ≈ +0.121 on average), which reflects variance redistribution across rotated PCA axes rather than improved separability. This effect is negligible on broader spectral ranges, highlighting the localization and instability of augmentation effects under small-n conditions. Overall, augmentation should be treated as an experiment on data geometry and model stability, and interpreted only together with strict baseline comparison, synthetic QC, and sanity checks—especially in small clinical datasets.