Assessment of blood serum stability with Raman spectroscopy and explanatory AI

This study explores the potential of conventional Raman spectroscopy and commonly used spectral analysis pipelines for rapid and straightforward assessment of degradation in serum samples resulting from storage delays. Serum samples from 18 volunteers were processed within 2 h of extraction, which later on were analyzed via Raman spectroscopy over 4 days, while the corresponding serum vials were kept at room temperature. The resulting spectra were processed, including silicon normalization and a newly proposed outlier detection ensemble method. Next, baseline correction was performed, and spectral unmixing along with Principal Component Analysis (PCA) were applied. Several classification models (KNN, RF, and SVM) were trained and evaluated on three distinct balanced datasets: one including all data, one excluding low signal-to-noise ratio (SNR) data, and one excluding low-SNR data with baseline correction. Feature importance, assessed through random permutations, was used for explainability. Spectral unmixing and PCA indicated limited spectral changes directly attributable to analyte degradation, with inter- and intra-sample variability dominating. Classification results showed that while removing the baseline led to inconclusive results, models trained on datasets retaining the baseline effectively identified non-degraded samples. These findings suggest that while conventional Raman spectroscopy may not be optimally sensitive to subtle analyte variations in serum stored at room temperature, the auto-fluorescence background holds promise as a potential biomarker for monitoring serum storage quality.