Machine Learning Interpretation of Optical Spectroscopy Using Peak-Enhanced Logistic Regression

Optical spectroscopy, a non-invasive molecular sensing technique, offers valuable insights into analyte composition, leading to advancements in material characterization, chemical identification, and disease diagnosis. Despite the informativeness, precision, and versatility inherent in high-dimensional optical spectra, their interpretation remains challenging. Machine learning methods have gained prominence in spectral analyses, efficiently unveiling analyte compositions. While achieving remarkable classification accuracy, these methods still face challenges in interpretability due to a lack of optimization for spectroscopy, large feature noise, and model complexity.<br/>In this context, we introduce a machine learning algorithm—logistic regression with peak-enhanced regularization (PE-LR)—tailored for spectral analysis. PE-LR demonstrates superior performance in both classification and interpretability, achieving an F1-score of 0.95 and a feature sensitivity of 1.0. This outperforms other methods such as elastic net logistic regression (E-LR), support vector machine (SVM), XGBoost, and principal component analysis followed by linear discriminate analysis (PCA-LDA).<br/>Our work highlights PE-LR as a robust tool for advancing the field of spectral analysis, offering enhanced interpretability and classification accuracy essential for studying complex analytes. Applying PE-LR to Raman spectra of Alzheimer's disease (AD) brain slices and healthy controls, we successfully identified AD brains and discerned spectral peaks distinguishing AD samples from healthy ones, suggesting potential disease biomarkers. Our findings underscore PE-LR's potency and promise in spectral analysis, with the capability to detect subtle spectral features and deliver precise, informative interpretations. Moreover, its adaptability extends seamlessly to other spectroscopic methods, including nuclear magnetic resonance (NMR), mass, and electron spin resonance (ESR) spectroscopy.