MAPPING MYOCARDIAL ACTIVATION DISTRIBUTIONS USING NEURAL NETWORKS - 2-D SIMULATION RESULTS

The goal of this study was to explore the capabilities of neural netwo rks to map with accuracy the sequence and location of myocardial activ ation using QRS complexes simulating normal and altered activation. A two-dimensional (2-D) fractal-based computer model of myocardial activ ation was used to develop training data for initial network learning. Two types of activation scenarios were used to evaluate network learni ng: 1) 450 training sets based on three activation foci per set using randomly chosen times and activation sites, and 2) 199 training sets b ased on a sequential, hierarchical blocking of the fractal-based model conduction network. Network learning was evaluated with training and test cases using trained weights. Network-calculated activation maps c ompared with the target activation maps had a mean error of < 5% in as signing the site and timing of activation. Pointwise mean correlation coefficients were > 0.98 for all conduction network cases and >0.84 fo r the more demanding point foci cases. We conclude, based on these sim ulation results, that neural networks may be used to calculate activat ion maps using electrocardiogram lead data for a variety of activation patterns.