Determining Heroin Geographic Origins: A Confidence-Driven LIBS Classification Method

In fifteen words or less, explain the significance of this contribution (Novel Aspect).: Enhances reliability of spectroscopic classification by propagating measurement uncertainties

Abstract Text:

While uncertainty is inherent in chemical measurements, current chemometric classification models rarely account for measurement uncertainty in their outputs. In spectroscopy, measurement uncertainty manifests as spectral noise, yet classification models don't typically translate it into classification confidence. We present a novel approach to estimate uncertainty in classification models using spectroscopic data. Our method generates synthetic noise samples based on replicate measurements' covariance matrix through Monte Carlo simulations, projects the uncertainty through the classification model, and estimates confidence intervals for predictions. When these intervals cross decision boundaries, samples are flagged as uncertain classifications. This approach bridges the gap between measurement uncertainty and classification confidence, providing more reliable results in spectroscopic analysis. This approach is nearly universal in application and shows how errors propagate differently through different models, e.g., PLS-DA vs. SVM with a Radial basis function. The approach was used to determine the confidence of classification of heroin origin following analyses by Laser Induced Breakdown Spectroscopy (LIBS). A hierarchical model was constructed to classify heroin based on 4 geographic regions – two in the Americas and two in Asia. The error propagation method shows the degree of granularity with which each sample can be classified with a determined confidence level.