SURVEY OF PATTERN RECOGNITION

One of the important uses of computers in clinical medicine is for the classification or screening of data. Examples abound where high speed and inexpensive but reliable automatic classification is desired. Electrocardiographs must be classified as healthy or abnormal. Differential white blood cell counts require the ability to discriminate between the various types of cells. Cancerdetecting smears must be sorted as normal or abnormal. It has become commonplace to speak of these kinds of sorting tasks as pattern-recognition problems and to advocate the application of pattern-recognition techniques for their solution. A wide range of such techniques exists. Some of them have been used by statisticians for years; others have been developed only recently as a result of the availability of high-speed computers. In this paper I shall describe some of the more common pattern-recognition methods. Although I shall cite some clinical applications of pattern recognition as illustrative examples, it is not my purpose to report on these in detail. An extensive literature'" already exists that provides numerous examples of the successful use of these methods. Indeed, many of the papers in this monograph will report the results of automatic classification experiments. I only hope that I can explain in this paper some of the unifying ideas that underlie many of the pattern-recognition methods already being applied. (See also a very good introductory article by Rosen.*) In order to apply pattern-recognition techniques, the phenomenon to be classified must be represented in some "computer-acceptable" form. Furthermore, the representation method used depends critically on the type of phenomenon. Thus, for photomicrographs of chromosomes, we might use complex picture-processing methods to represent the picture as a list of numbers, whereas, for a medical history record, it may only be possible to represent the data on the form as a list of nonnumerical symbols. The phenomenon to be classified is called the event and its representation is called the data in order to distinguish between them. In this paper, I shall be primarily concerned with those pattern-recognition techniques for sorting numerical data, that is, representations that are in the form of a list of numbers. First, I shall describe several data-classifying methods (assuming that the event