Date of Graduation


Document Type


Degree Type



Eberly College of Arts and Sciences


Forensic and Investigative Science

Committee Chair

Glen P. Jackson

Committee Co-Chair

Jacqueline Speir

Committee Member

Jacqueline Speir

Committee Member

Kenneth Ryan


Fentanyl is an opioid that is about 100 times more potent than morphine, and its analogs are just as potent, and some like carfentanil are even more so potent. In recent years, fentanyl and its analogs have been responsible for over 50% of overdose-related deaths. In response, the DEA has placed all fentanyl analogs as Schedule I compounds because of their non-medicinal use and high potential for abuse and physical dependence. However, fentanyl analogs have high structural and mass spectral similarities due to the extremely conservative nature of the core fentanyl structure, which makes the differentiation and identification of fentanyl analogs by conventional algorithms difficult. Seized drugs analysts typically use a consensus approach to compare an unknown spectrum to an ideal spectrum from their database. Still, novel fentanyl analogs are emerging faster than databases can be updated and purchasing standard reference materials are very costly. Seized drug analysts can benefit from an Expert Algorithm for Substance ID (EASI) algorithm that would improve the confidence of drug identification and enable inter-laboratory identifications hence obviating the need for acquiring concomitant spectra of standards. A database containing over20,000 replicate electron-ionization (EI) mass spectra of nine fentanyl analogs was compiled from two laboratories. Each fentanyl analog was used as a known positive (KP) for model building, while the remaining eight served as corresponding known negatives (KN). The 20 most abundant ions were extracted, and the abbreviated spectra were randomly divided into training and testing sets. Twenty generalized linear regression models (GLM) were built for each compound by sequentially using the abundance of each ion as the dependent variable and the abundance of the 19 remaining ions as the independent variables. The independent variables were entered stepwise until there were no significant changes to the predicted models. The predicted abundances were compared to the measured abundances using a variety of similarity metrics like the Pearson product-moment correlation (PPMC) and dissimilarity metrics like mean absolute residual and Euclidean distance. Each metric of scoring was used as a binary classifier to determine true positive (TP), true negative (TN), false positive (FP), and false negative (FN) rates over a range of threshold values. These classifications were then used to plot a receiver operating characteristic (ROC)curve to calculate the area under the curve (AUC).On average, EASI outperformed the consensus approach in every metric. The residuals in the predictions for the known positives typically improved by a factor of 3 over the consensus approach. The ranges of the EASI Euclidean distances for the known positives were consistently smaller than the consensus approach. The PPMC values between the measured and predicted spectra of known positives of the selected drug models exceeded 0.9666 for the training and testing sets. Known negatives in the validation set typically had smaller PPMC values than the smallest PPMCs for known positives, resulting in AUCs no less than.960in the ROC plots for binary classification.

Embargo Reason

Publication Pending