In this study myocardial infarction was localised by a Learning Vector Quantization (LVQ) classifier. Only information about ST-elevations in all 12 leads of the standard ECG were used. The significance of proper initialisation is demonstrated. A total classification accuracy of 85.6% was achieved by a classifier trained with the optimized-learning rate LVQ1 and 50% of the 769 patients. When the classifier was further trained with the LVQ2.1 and the LVQ3 algorithms no significant improvement in the classification accuracy was observed.