Improving LIBS machine learning bacteria classification performance based on self-absorption correction

Abstract

Rapid and accurate bacterial detection is vital for the effective management and mitigation of risks associated with bacterial contamination. Recently, the integration of laser-induced breakdown spectroscopy (LIBS) and machine learning provides an effective pathway for the rapid identification of bacteria. However, the ubiquitous self-absorption effect in LIBS results in non-linear attenuation of spectral intensity and distortion of peak profiles, which introduces systematic errors, diminishes the physical consistency of spectral features, and limits the discriminative performance of machine learning classifiers for high-concentration samples. To address these issues, this study presents a self-absorption correction method based on the ratio of spectral line intensities and self-absorption parameters to improve the classification performance of bacterial LIBS spectra. The method utilizes elemental doublets with similar upper and lower energy level structures to retrieve self-absorption coefficients through the relationship between experimental intensity ratios and theoretical parameter ratios, which achieves effective intensity correction without direct measurement of complex plasma parameters. To systematically evaluate the effectiveness of this correction, spectral data before and after correction are processed using four standard machine learning models, specifically partial least squares discriminant analysis (PLS-DA), principal component analysis combined with k-nearest neighbor (PCA-KNN), support vector machine (SVM), and random forest (RF), to classify seven bacterial species. Results indicate that self-absorption correction significantly enhances classification performance across all models, with the overall accuracies of PLS-DA, PCA-KNN, SVM, and RF increasing from 76.8%, 92.7%, 95.9%, and 95.9% to 87.0%, 99.2%, 98.0%, and 98.8%, respectively. Key evaluation metrics, including area under the curve (AUC), average precision (AP), Precision, Recall, and F1-score, also show systematic improvements, where PCA-KNN and RF approach ideal classification levels across multiple indicators. In conclusion, the proposed dual-line intensity ratio correction method effectively suppresses self-absorption in LIBS spectra and substantially improves the accuracy and robustness of machine learning models for bacterial classification. This approach provides technical support for constructing high-precision LIBS systems and possesses potential applications in food safety, clinical diagnosis, and biosafety monitoring.