Estimating Probabilistic Confidence for Mixture Components Identified Using a Spectral Search Algorithm

Providing a confidence measure associated with the substance(s) identified in an unknown mixture by a spectral search technique is critical for non-expert users of devices and techniques based on spectroscopy. In this work, a technique for estimating probabilities associated with substances identified by spectral searching is described. In the proposed approach, a mixture analysis algorithm processes the spectrum of an unknown sample using a spectral library to generate a list of substances that may be present in the sample. The partial correlation of each of the substances in the list is then computed. The estimation of the probability is accomplished through a generalized linear model that converts the partial correlation values to a probability measure for each of the mixture components. The statistical properties of partial correlation allow probability estimation irrespective of whether a substance is present in a pure form or within a mixture. The technique was evaluated using both simulated and real Raman spectra of multi-component mixtures, and adequate performance was demonstrated.