A Combined Statistical and Neural Network Approach for Prediction of Relapse in Patients with Chronic Alcoholism using QEEG

This paper describes the application of quantitative EEG for the discrimination of relapse and abstinent chronic alcoholics using data from 29 abstainers and 49 relapsers, all of them are unmedicated chronic alcoholics. Electrode positions were used according to the 10/20 system. One representative 30 s artefact-free EEG-segment was analysed of each patient. For quantitative analysis (pre-processing) we chose spectral as well as Hjorth’s parameters and the correlation dimension of Grossberger-Proccacia. Spectral values: Power within the bands Delta (0.5-5.5Hz), Theta (6.0-8.0Hz), Alpha (8.5-12Hz), Beta1 (12.5-18Hz), Beta3 (21-30Hz) for all 19 electrodes. Hjorth’s parameter: Mobility (mean frequency), variability of activity (amplitude), complexity (frequency) were used for the bipolar pairs: F3-F7, F4-F8, P3-O1, P4-O2, T3-P3, T4-P4 The neural network approach uses a feed forward net with one hidden layer. The formal criteria for this approach were not well suited because of the large number of potential network inputs compared to the number of data-sets to be used for training. Therefore various selection schema’s for the network inputs had been applied to increase the generalisation ability on the independent test-set: Subsets of electrodes, frequency bands and aperiodic features (Hjorth’s parameters, correlation dimension) had been chosen based on medical considerations. Furthermore, genetic algorithms were used to build sub-optimal input sets and network topologies regarding the generalisation. Finally a input pruning strategy was applied based on a global input sensitivity measure of the trained net. Comparison with statistics on the identical training/test-sets uses approximated F-test for Wilk’s lambda with random and stepwise introduction of variables. Best results using a 5 variable model resulted in an overall correct classification of 75% (sensitivity: 96%, specifity: 40%). The direct neural network approach (medical motivated input selection) increased the overall correct classification rate to 80% for the independent test-set but with a different sensitivity/specifity relation. The additional application of genetic algorithms and neural sensitivity analysis (pruning) helps to reduce the problems of selecting useful network inputs. Thus, a further improvement of the classification results was possible up to 85% (overall), although the required computing power for this approach was considerably higher than for the linear statistics.

Zurück zur Hompage der Cowan!