Machine Learning Interpretation of Optical Spectroscopy Using Peak-Sensitive Logistic Regression.

Area of Science:

• Spectroscopy and Spectrometry
• Machine Learning Applications
• Materials Science

Background:

• Optical spectroscopy provides rich data for material and biosample analysis but faces interpretation challenges.
• Machine learning (ML) aids spectral analysis but often lacks clear feature importance maps.
• Existing ML methods struggle with spectral noise, model complexity, and spectroscopy-specific optimization.

Purpose of the Study:

• Introduce a novel ML algorithm, logistic regression with peak-sensitive elastic-net regularization (PSE-LR), for enhanced spectral analysis.
• Improve classification accuracy and interpretability in optical spectroscopy.
• Develop a method that generates peak-sensitive feature importance maps for spectral data.

Main Methods:

• Developed PSE-LR, a logistic regression model with peak-sensitive elastic-net regularization.
• Applied PSE-LR to analyze Raman and photoluminescence (PL) spectra.
• Compared PSE-LR performance against KNN, E-LR, SVM, PCA-LDA, XGBoost, and NN.

Main Results:

• Achieved an F1-score of 0.93 and a feature sensitivity of 1.0 with PSE-LR.
• Successfully detected SARS-CoV-2 spike protein's RBD at ultralow concentrations.
• Identified neuroprotective solution (NPS) in brain samples and characterized WS2/WSe2 heterostructures.
• Analyzed Alzheimer's disease (AD) brains and identified potential biomarkers.

Conclusions:

• PSE-LR effectively detects subtle spectral features and generates interpretable importance maps.
• The algorithm is beneficial for material, molecule, and biosample characterization using spectroscopy.
• PSE-LR facilitates the development of advanced nanodevices like nanosensors and miniaturized spectrometers.