Mass spectral search method using the neural network approach

This paper explores the use of neural networks as a novel approach in the implementation of spectral library search for gas chromatography mass spectrometry which is a common and powerful analytic tool for the forensic drug chemists nowadays. A total of 28 drugs currently under control in Hong Kong were chosen for the study. Real forensic data, which represents mass spectra obtained under various conditions ranging from good to poor, were used for training and testing. Salient features from the spectral data were extracted using Simpsonʹs diversity index, and such features were presented to various neural networks as input. A total of 355 spectra were used for training the neural networks, and a further set of 163 spectra was used for evaluation. All the neural networks performed well, with recognition rates above 97.5%. Moreover, the best performing neural network achieved perfect recognition.