Preconditioned K-SVD for ECG Anomaly Detection

We propose Preconditioned K-SVD, an algorithm for dictionary learning adapted for sparse coding with Least Squares Pursuit. The algorithm resembles K-SVD iterations but operates in a preconditioned signal space, which implies a different optimization goal then K-SVD, namely to minimize the norm of the cosupport coefficients in the sparse representations. The algorithm is designed to operate with sparse coding algorithms based on null space projections, e.g. Least Squares Pursuit. Simulation results in the context of ECG arrhythmia detection show a performance similar to K-SVD with Orthogonal Matching Pursuit sparse coding.