Robust and interpretable SERS classification via a machine-learning evaluation framework

Yuting Liu

doi:10.1016/j.vibspec.2026.103948

Robust and interpretable SERS classification via a machine-learning evaluation framework

2026-07-17

0

PRE

AI

DOI：10.1016/j.vibspec.2026.103948

Original

Original request for help

Save

Abstract

En 中文

Surface-enhanced Raman spectroscopy (SERS) enables rapid, label-free bacterial fingerprinting, yet many machine-learning studies lack transparent evaluation protocols and interpretable feature analysis. This study presents a SERS-machine-learning evaluation framework for bacterial classification using a fixed-concentration ten-class Staphylococcus aureus dataset (100 spectra) and leak-free 60/20/20 train/validation/test partitions repeated across 15 random seeds. After Savitzky-Golay smoothing and standard normal variate normalization, spectra were analyzed in both full (384 variables) and cropped (400–1200 cm⁻¹; 298 variables) representations. Linear discriminant analysis (LDA) served as the primary dense classifier, while stability-selection-based feature selection combined with LDA (StabSel-LDA) was evaluated as a sparse companion model with validation-tuned feature count K. LDA achieved near-ceiling clean performance (mean accuracy 0.970–0.973; macro-F1 0.966–0.969) and remained stable under additive noise, drift, and ±5 cm⁻¹ spectral shifts (accuracy 0.953). StabSel-LDA reduced the feature set to 124–228 bands and produced reproducible band-selection profiles across seeds, but showed lower clean accuracy (0.940–0.943) and pronounced degradation under ±5 cm⁻¹ shifts (0.760–0.783). These results support a role-separated deployment strategy in which LDA functions as the primary decision model, while StabSel-LDA provides a sparse companion for feature compression and interpretable spectral-band identification. The proposed framework offers a transparent approach for balancing classification performance, robustness assessment, and spectral interpretability in small SERS datasets.

Keywords:

Surface-enhanced Raman spectroscopy

Bacterial strain classification

Chemometrics

Sparse modeling

Stability selection

Analytical methods

Robust and interpretable SERS classification via a machine-learning evaluation framework

Abstract

Journal

Organization

Cited Papers