Bayes' theorem applied to chromosome classification

Abstract

As part of an interactive, computer-aided image processing system developed at Tufts-New England Medical Center, a computer program to classify human chromosomes into ten major groups has recently been devised. The program applies Bayes' theorem to rankings to determine the most probable group into which a chromosome will fall. A method of calculating the required distributions from a relatively small sample is given. The program then proceeds with additional constraints to construct a karyotype. Four parameters derived from the measurements are used: normalized length, normalized area, and centromeric indices by length and area. On a sample of 13 800 chromosomes from 300 “normal” spreads, there were 1.3% discrepancies between manually determined and computed group assignments. Two hundred and nineteen out of the 300 spreads had no discrepancies. Pair assignments within the ten major groups were also computed. Agreement with manual assignments was relatively low but much better than chance, as shown by quantitative estimates. The computed method appears to be similar to human performance, quantitatively justifying karyotyping into the ten groups but pointing out the need to develop new methods for identifying pairs other than 1, 2, 3, and 16. The adaptations needed to make Baye's theorem applicable to this problem may extend its usefulness to other biomedical areas.