Feature-extracted SERS combined with enhanced machine learning and deep learning for fast screening of COVID-19...

Surface-enhanced Raman spectroscopy (SERS) has emerged as a powerful label-free technique for molecular-level identification of infectious agents. This study establishes an approach through label-free feature-extracted SERS integrated with enhanced machine learning (ML) and deep learning (DL) algorithms for fast screening of COVID-19 in inpatient specimens. SERS-derived SARS-CoV-2 signature spectra were used to classify inpatient specimens as CoV (+) or CoV (-) using Au NPs/fZrO₂ and Au NPs/pZrO₂ substrates. All spectra underwent standard preprocessing, including cosmic spike removal, baseline correction, and Savitzky-Golay smoothing, and statistical outlier elimination based on Hotelling's T² and Q-residual analysis. Multiple normalization strategies (none, min-max, and z-score) and feature extraction approaches (full spectra, manual peak selection, and principal component analysis (PCA)) were evaluated. Support vector machine (SVM) classifiers were trained using k-fold cross-validation and subsequently assessed via blind testing on independent inpatient samples. Although z-score normalization combined with PCA yielded the highest validation accuracy (>98%), min-max normalization with manual peak selection demonstrated superior generalization, achieving blind test accuracies of 87.5% and 75% for Au NPs/fZrO₂ and Au NPs/pZrO₂ substrates, respectively. To further enhance classification performance, artificial neural network (ANN)-based DL models were implemented, achieving optimal blind test accuracies of 93.75% for Au NPs/pZrO₂ and 87.50% for Au NPs/fZrO₂. Appropriate data processing significantly improves classification accuracy. These findings indicate that effective SERS-based detection relies on coordinated optimization of substrate design, data normalization, feature selection, and classification models, enabling robust early screening of infectious diseases such as SARS-CoV-2.