Machine learning methods for classifying novel fentanyl analogs from Raman spectra of pure compounds

In previous research, we demonstrated the promise of detecting novel fentanyl analogs from mass spectra using machine learning models. This approach complements existing library matching methods and provides a key capability amid the recent sharp increase in abuse of fentanyl and its analogs. However, many applications rely upon portable devices such as Raman spectrometers, rather than mass spectrometers that are generally located in laboratories. In response, we adapted our models to Raman-based sensing, devising a machine learning approach for detecting novel fentanyl analogs from Raman spectra. Whereas mass spectra consist of well-defined discrete peaks, Raman spectra are continuous. To aid model development, we extracted features from each spectrum using smoothing, background subtraction, and principal component analysis (PCA). Additionally, we extracted features related to spectral peaks and similarity to spectra of known compounds; these features were guided by subject-matter expertise. We also used a third feature set that combined the features from PCA and from spectral peaks. With these three feature sets as inputs, we developed fentanyl analog classification models using various machine learning techniques. These techniques included multi-layer perceptron, neural network, partial least squares, penalized multinomial regression, random forest, regularized logistic regression, support vector machines, and extreme gradient boosting. We developed and tested our models using 320 Raman spectra of pure compounds, assessing performance via cross-validation. Models with PCA-based features performed better than those using expert-based features, achieving more than 90% probability of detection alongside less than 1% probability of false alarm. Although fentanyl compounds are often found with other components (e.g., cutting agents), some applications such as mail screening may encounter relatively pure fentanyl analogs. The results within this study suggest that our machine learning models are particularly promising for such applications.