Pattern recognition applications

G. N. Stephanopoulos, Knowledge-based systems, artifidal neural networks and pattern recognition applications to biotechnological processes, Curr. Opin. Biotechnol. 1996, 7, 231-234. 

ARTHUR Infometrix, 2200 Sixth Ave. 833, Seattle, Wash. 98121, USA cca 7000. This package of Fortran programs has been developed at the Department of Chemistry of the University of Washington, Seattle. For many years it was the most widely used program for PCA and other pattern recognition applications in chemistry. ARTHUR has been written for mainframe computers, but also PC-versions are used (e.g. TNO CIVO Institute, P.O.Box 360, NL-3700 AJ Zeist, The Netherlands). An overview of ARTHUR is given in Wolff and Parsons (ref. 14). 

The results of many pattern recognition applications in chemistry are doubtful because an important prerequisite of the data was not fulfiI led ... 

The general concept of pattern recognition is applicable to various classification problems in science and technology. Table 1 gives a short list of current pattern recognition applications. 

An evident quality criterion for a classifier seems to be the percentage of correctly classified patterns (overall predictive ability). This criterion was used during the first years of pattern recognition applications in chemistry. However, the overall predictive ability suffers from the fact that it depends extremely on the composition of the prediction set. If, for example, 90 % of the patterns belong to class 1 and 10 % to class 2, a primitive "classifier" that always predicts class 1 would have 90 % overall predictive ability. Therefore, overall predictive ability has to be refused if an objective characterization of classifiers is necessary. 

Due to the large number of papers dealing with pattern recognition applications in mass spectrometry it may be useful to list some reviews on this topic (alphabetically). 

TABLE 10. Pattern recognition applications in mass spectrometry with specialized data sets. 

One of the first pattern recognition applications in mass spectrometry was the attempt to determine the molecular formula by a decision tree C120, 128, 1293. The decision tree contained several binary classifiers. Each of the classifiers decided whether a compound contains more atoms than a given number- A run through the decision tree yields the molecular formula of an unknown whose low resolution mass spectrum is known. A tree with 26 classifiers was necessary for a set of 346 compounds of formulas --i 6 0-3 0-2 spectra with an artifi-... 

The most chal lenging task of pattern recognition applications in mass spectrometry is the automat i c recogni t i on of molecular subst ruetures in a molecule. In a g reat number of papers c Lassi fi ers for this purpose were developed by the learning machine method typical predictive abilities are 70 to 95 %. 

Short reviews about pattern recognition applications in infrared spectroscopy have been given by Isenhour and Jurs C1173, and Kowalski C1483. 

Only a few works have been performed on pattern recognition applications in Raman spectroscopy. This is probably due to the lack of suitable computer-readable spectra. Methods of coding and classification would be the same as used for infrared spectra. 

Pattern recognition is a powerful tool in the identification of archaeological artefacts. A typical example of a pattern recognition application in chemistry is the classification of obsidian artefacts by Kowalski et. al. C1623. A total of 45 obsidian samples from different sources in northern California and 27 archaeological obsidian artefacts of unknown origin were analyzed by X-ray fluorescence spectroscopy. 

The most difficult problem in pattern recognition applications to SAR-classifications is the formulation of meaningful descriptors that describe the molecular structure and are correlated with the classification problem. The widely used concept of a linear, binary classifier assumes a linear relationship between the structural properties (pattern components) x. and the biological activity. 

Further pattern recognition applications deal with serum chemistry C210, 3283, treatment of arthritis C3313, diagnosis of cancer C303, shock C3383, and myocardial infarction C383. 

Chapters 9 and 10 deal with preprocessing of original data and feature selection. These problems must be treated at the beginning of a pattern recognition application. These Chapters have not been positioned at the beginning of the text because a more detailed description of these subjects requires some basic knowledge of pattern recognition methods. 

Chapter 12 gives an overview about pattern recognition applications in chemistry. Chapters 13 to 20 extensively describe applications in spectral analysis, chromatography, electrochemistry, material classification, Structure-activity-relationship research, clinical chemistry, environmental chemistry and classification of analytical methods. 

Pattern Recognition Applications to Large Data-Set Problems, Sing-Tze Bow... 

---