Neural approaches to the diagnosis and characterization of the Lyme disease

Our primary goal was to develop and to study neural approaches to classification and rule extraction. We used the Lyme database for the necessary validation phases, the database was recorded to study the Lyme borreliosis, by the Italian Working Group on Lyme disease. A procedure using a combination of statistical analysis and a standard supervised neural method obtained a very good classification (diagnosis) performance, but the resulting rules (characterization) were readable only in a rough form. A different approach was used to extract more readable rules. A binary-units network was trained by an information-theoretic learning rule (CCE). The result is a hierarchy of rules describing the classification by a sequence of yes/no questions, confirming the known diagnostic criteria to a satisfactory extent. The next step was to obtain confirmation from an unsupervised method. An auto-associative network trained by backpropagation was used. The output of this method was a clustering of the patterns, and their interpretation led to the same conclusions as the CCE approach