The following is an abstract for the selected article. A PDF download of the full text of this article is available here. Members may download full texts at no charge. Non-members may be charged a small fee for certain articles.

Development of a Hierarchical Classification System with Artificial Neural Networks and FT-IR Spectra for the Identification of Bacteria

Volume 54, Number 10 (Oct. 2000) Page 1471-1479

Udelhoven, Thomas; Naumann, Dieter; Schmitt, Jürgen

The practical value of elaborated vibrational spectroscopic techniques in medical and microbiological biodiagnostics depends strongly on the reliability, the speed, the ease of use, and the evaluation procedures of the acquired data. In the present study, artificial neural networks (ANNs) were used to establish a hierarchical classification system for microbial Fourier transform infrared (FT-IR) spectra suitable for identification purposes in a routine microbiological laboratory. A radial basis function network (RBF) proved to be superior for a top-level classification of the FT-IR spectra at the genus level. Species within these genera were sequentially further classified by using multilayer perceptrons (MLPs), which achieved a larger differentiation depth than RBF networks. The MLPs were trained with several learning algorithms. Best performance was achieved with the cascade correlation (CC) approach to determine the network topology combined with resilient propagation (Rprop) as the training algorithm. The final hierarchically organized model was able to discriminate between four genera of microorganisms comprising 42 different strains of Pseudomonacae, 33 strains of Bacillus, 46 strains of Staphylococcus, and 6 species and 24 strains of yeast genera Candida. Altogether, 145 strains from international microbial strain collections are comprised in 971 spectra. The species Candida albicans could be further classified with respect to susceptibility against the antibiotic drug fluconazole, which is of therapeutic relevance. Key factors for the classification results of the bacterial FT-IR spectra were the data pretreatment, the number of wavelengths selected by a feature extraction algorithm, the type of network, and the learning function used for the ANN training.