Computerized classification of congenital malformations using a modified Bayesian approach.

The diagnostic classification of children with dysmorphic features involves over 200 syndromes and 232 findings, with an average of about 15 findings per syndrome. A knowledge base expressed in terms of Boolean combinations of findings is impractical. The normal Bayesian method requires a very large incidence matrix with the vast majority of cells being zero. A modified Bayesian method is proposed in which each syndrome is described in terms of its associated findings, whose incidence P (S/D) are designated as essential (0.90), prevalent (0.90), occasional (0.70) or rare (0.15), whilst P(S/-D) ranged from (0.08) to (0.10). The Bayesian calculation determines the probability of the presence P(D/S) or the absence P(-D/S) of each syndrome. The differential diagnosis consisted of all syndromes whose presence has a probability greater than 0.85. One hundred and thirty-one cases from the Hanna Khoushi Developmental Pediatrics Center at Haifa's Rothschild Hospital were considered. Of the 42 cases for which the center's specialists reached a diagnosis, the system listed the correct diagnosis for 91%. The system reached a diagnosis in about half of the remaining 89 cases. The medical literature is arranged by syndrome whilst the computer allows a case by case approach, thereby avoiding the need for the physician to consider each syndrome to see if it fits his case. This study shows that our modified Bayesian analysis is a valid method for shortening the physician's search in an area of great diagnostic complexity.