In this work, an automated procedure for extracting chemical profiles of illicit drugs from chromatographic-mass spectrometric data is presented along with a method for comparison of the profiles using Bayesian inference. The described methods aim to ease the work of a forensic chemist who is tasked with comparing two samples of a drug, such as amphetamine, and delivering an answer to a question of the form 'Are these two samples from the same source?' Additionally, more statistical rigour is introduced to the process of comparison.
The chemical profiles consist of the relative amounts of certain impurities present in seized drug samples. In order to obtain such profiles, the amounts of the target compounds must be recovered from chromatographic-mass spectrometric measurements, which amounts to searching the raw signals for peaks corresponding to the targets. The areas of these peaks must then be integrated and normalized by the sum of all target peak areas.
The automated impurity profile extraction presented in this thesis works by first filtering the data corresponding to a sample, which includes discarding irrelevant parts of the raw data, estimating and removing signal baseline using the asymmetrical reweighed penalized least squares (arPLS) algorithm, and smoothing the relevant signals using a Savitzky-Golay (SG) filter. The SG filter is also used to estimate signal derivatives. These derivatives are used in the next step to detect signal peaks from which parameters are estimated for an exponential-Gaussian hybrid peak model. The signal is reconstructed using the estimated model peaks and optimal parameters are found by fitting the reconstructed signal to the measurements via non-linear least squares methods. In the last step, impurity profiles are extracted by integrating the areas of the optimized models for target compound peaks. These areas are then normalized by their sum to obtain relative amounts of the substances.
In order to separate the peaks from noise, a model for noise dependency on signal level was fitted to replicate measurements of amphetamine quality control samples non-parametrically. This model was used to compute detection limits based on estimated baseline of the signals.
Finally, the classical Pearson correlation based comparison method for these impurity profiles was compared to two Bayesian methods, the Bayes factor (BF) and the predictive agreement(PA). The Bayesian methods used a probabilistic model assuming normally distributed values with normal-gamma prior distribution for the mean and precision parameters. These methods were compared using simulation tests and application to 90 samples of seized amphetamine.
