A pattern-recognition approach for lead-selection in heartbeat detection

In this work, we developed and evaluated an algorithm for selecting the most suitable lead for performing heartbeat detection in ECG signals. For the development and evaluation we used a public dataset of 927 multilead (2-12 leads) stress-test recordings, with manually reviewed heartbeat locations. The algorithm consists of a pattern-recognition block based on features calculated from the RR interval sequence, and a mixture of Gaussian classifier. This block estimates whether the heartbeat is correctly detected, ommited or incorrectly detected. With these estimations, a detection quality index is calculated from the sensitivity (S) and positive predictive value (P⁺) of each lead. With this quality index a decision is made to choose the best lead. Results show that the correct lead has been selected in 70% of the recordings, and in 93% of the recordings the best lead was among the top 3 leads with higher detection quality index. Finally, the selection of the lead with higher quality index produces a gross median S of 100%, with percentile 5 at 99.6, and a gross median P⁺ of 98.9%, with percentile 5 at 89.2. The algorithm was developed and evaluated using ECG signals, but could be used with other cardiovascular signals as well, being suitable for automatically selecting the best lead/signal, or sorting them for further analysis or manual correction.