Self-organizing visualization and pattern matching of vectorcardiographic QRS waveforms

• Thisdevelops a new self-organizing method to visualize pattern similarity between 3-lead VCG signals.
• This work performs pattern matching of VCG signals to cluster subjects with similar space-time cardiac electrical activity.
• This work derives a large-scale network by representing each subject as a node and optimizing node locations to identify community structures.
• Self-organizing pattern matching shows strong potentials for large-scale unsupervised learning of patient groups.

QRS morphology is commonly used in the electrocardiographic diagnosis of ventricular depolarization such as left bundle branch block (LBBB) and ventricular septal infarction. We investigated whether pattern matching of QRS loops in the 3-dimensional vectorcardiogram (VCG) will improve the grouping of patients whose space-time electrical activity akin to each other, thereby assisting in clinical decision making. First, pattern dissimilarity of VCG QRS loops is qualitatively measured and characterized among patients, resulting in a 93×93 distance matrix of patient-to-patient dissimilarity. Each patient is then represented as a node in the network (or a star in the galaxy), but node locations are optimized to preserve the dissimilarity matrix. The optimization is achieved with a self-organizing algorithm that iteratively minimizes the network energy. Experimental results showed that patients’ locations converge as the representation error reaches a stable phase. The convergence is independent of initial locations of network nodes. Most importantly, 93 patients are automatically organized into 3 clusters of healthy control, LBBB, and infarction. Spatial coordinates of nodes (or patients) are evidently novel predictors that can be used in the computer-assisted detection of cardiac disorders. Self-organizing pattern matching is shown to have strong potentials for large-scale unsupervised learning of patient groups.