Classification of human chromosomes in context using constrained Markov networks and Bayesian probability

This thesis presents a system for the classification for human chromosomes in cell context, using constrained Markov networks and Bayesian probability. The system presupposes that the networks already exist, the chromosomes have been classified in isolation, a normal cell is analyzed such that there are no missing or extra chromosomes, and the sex chromosomes are excluded. Statistical methods, alignment probabilities from constrained Markov networks, and information when considering the chromosomes as isolated objects are used to classify chromosomes in cell-context.

For classification in cell-context, classification accuracy is 100% until the confusion matrix is changed. Since the channel model is the fundamental structure, the probability of correct classification in cell-context is much greater than the probability of any errors. The product of probabilities is close enough that the correct chromosome types are forced. As a result, the system always chooses the correct vector which maximizes overall classification. The results are surprising, but the performance is degraded by significantly altering the Classification Confusion Matrix. Classification accuracy improves by 3.1% for classes A and B combined and about 10.2% for classes F and G combined compared to previous results for classes A, B, F, and G in isolation.