Automatic recognition of handwritten medical forms for search engines

Handwriting recognition (HR) is a challenging problem that is made tractable only by the contextual constraints offered by specific applications. The population of a national emergency medical service database from the collection of the New York State (NYS) Pre-hospital Care Report (PCR) calls for handwriting recognition. Such a database can enable emergency preparedness, response, and homeland security. We address several research challenges presented by the task of reading hand-filled PCR's in particular and medical forms in general. Written text on such forms has poor legibility due to insufficient size of writing areas (e.g. compressed text or text curved along margin), vehicle motion, writing with gloves, and the immediacy of the emergency environment. Challenges include: (i) written matter often spilling beyond the form boundaries, (ii) diverse lexicons in the medical domain, and (iii) low recognition performance due to poor legibility of text. A fourth challenge is that modern search engines expect to operate on known text and not on handwriting. In order to address these issues, we have developed the following: (i) the first text extraction technique which operates on carbon paper, (ii) a lexicon reduction strategy which maps partial recognition information to medical topic categories, and (iii) an information retrieval system capable of searching forms using handwriting recognition results.; While the emphasis of this research is on medical forms, the ideas extend to any domain in which there is at least one sentence of text that can be classified under high level topic categories. In the application of medical forms, it is shown that the words written by health care professionals involved in all aspects of patient assessment can be organized within the context of anatomical positions. Conceivably, if a patient with a broken leg is rescued, then the handwriting will be related to the identification and rescue efforts involving the anatomical position of legs.; The primary issue is how the category is determined if the handwritten words are unknown. Since both the lexicon reduction step, aimed at improving the recognition performance, as well as the search engine require the recognized words, a new paradigm must be developed to solve the problem. The algorithm described in this research automatically learns the salient relationships between characters from correlated words, and maps such related characters to categories. (Abstract shortened by UMI.)...