QRS detectors performance comparison in public databases

Automatic QRS detection remains a challenging task in certain types of recordings, limiting the capacity of automating subsequent tasks that heavily depends on proper heartbeat location. Performance estimation of these algorithms is calculated almost exclusively in a few databases, ignoring the generalization to other more complex situations. In this work, we evaluated six QRS detection algorithms in 13 ECG databases. Four out of the six algorithms, and 11 out of 13 databases are publicly available. The databases were categorized into 5 groups: normal sinus rhythm, arrhythmia, ST and T morphology changes, stress-test and long-term. The best evaluated algorithm was gqrs, achieving S of 95 (85-98) (median and percentile range 5-95) and P+of 93 (90-96) across all databases. When analyzing the performance by groups of databases, this algorithm obtained the first rank in 4 out of 5 groups. The algorithm developed in our group achieved a performance close to gqrs, and obtained the best performance in the stress group. This evaluation setup includes a broad variety of recordings, being useful to estimate the actual performance of QRS detection algorithms, not only in a global sense but also specific to specific type of recordings.