Recognition of Diagnostically Useful ECG Recordings: Alert for Corrupted or Interchanged Leads Irena Jekova 1 , Vessela Krasteva 1 , Ivan Dotsinsky 1 , Ivaylo Christov 1 , Roger Abächerli 2 1 Institute of Biophysics and Biomedical Engineering, Bulgarian Academy of Sciences, Sofia,Bulgaria 2 Biomedical Research and Signal Processing, Schiller AG, Baar, Switzerland Abstract The upgrade of mobile phones with applications for acquisition, pre-processing and transmitting the patient’s ECG to a hospital unit would be of great benefit for prevention against the most frequent mortality caused by heart failure. This idea is promoted by the Computing in Cardiology Challenge 2011, which encourages the development of algorithms for analysis of the ECG quality within few seconds, aiming to warn about diagnostically unacceptable recordings. This paper presents an algorithm for scoring the noise corruption level by evaluation of ECG amplitude dynamics, baseline wander, powerline interference, EMG and peak artifacts. The score achieved for participation in Event1 is 0.908. Additionally unacceptable ECGs with interchanged leads are detected with sensitivity of 96.8% (30/31 files) for peripheral leads and 87% (40/46 files) for chest leads. 1. Introduction When conditions during ECG acquisition are not rigorously controlled, ECG quality is highly susceptible to external noisy components and other distorting factors which might impede the reliable manual or automated measurements, or hazard the correct diagnosis. Automatic management of large amount of ECGs by analytical quality metrics is shown to improve the quality of ECG annotations reducing human review and costs [1,2]. During the years, members of our team are contributing towards development of methods for improving the ECG quality by filtering the main sources for ECG corruption - powerline interference (PLI), baseline wander (BLW) and electromyographic (EMG) noise. The main goal is to maximally preserve the useful ECG components, commonly overlapped with noises. In this respect, the subtraction procedure eliminates PLI with amplitude and frequency deviation without affecting the ECG spectrum [3]; the BLW bi-directional high-pass recursive filter [4] is optimized towards adapting the cut- off frequency with respect to the frequency components of the ECG signal [5]; the approximation filtering with dynamically varied number of samples and weighting coefficients in respect to the ECG slope, is preserving sharp QRS forms with a considerable reduction of the EMG noise [6]; the ‘linearly-angular’ procedure for EMG suppression is applying smoothing filtration outside the QRS complexes, and moving averaging inside them with restoration of the sharp Q, R and S peaks [7]. Misplacement of electrodes in 12-lead ECG is reported in 0.4-4% of all clinical recordings – a severe cause of erroneous diagnosis due to simulated false or concealed true ECG abnormalities [8]. Batchvarov et al [9] review the effect of the most common cases for interchange in peripheral and chest leads on P-QRS-T patterns, together with some algorithms for their detection. Specific cable interchanges or ECG abnormalities might disturb the correct detection. The presented method detects noise corruption and leads interchange for recognition of diagnostically useful ECGs in the Computing in Cardiology Challenge 2011. 2. ECG dataset The study uses the Challenge 2011 dataset available from PhysioNet [10], including 10-second recordings of standard 12-lead ECGs (sampled at 500Hz, 5µV/LSB resolution, full diagnostic bandwidth 0.05–100Hz). The dataset comprise signals related to common problems which might appear when people with varying amounts of training are recording ECG via disposable or suction cup electrodes connected to mobile phones (misplaced electrodes, poor skin-electrode contact, not connected electrode, PL interference, artifact resulting from patient motion, etc.). Reference annotations of the ECG quality in the context of ‘acceptable’ or ‘unacceptable’ recording for diagnostic interpretation are accessible for the challenge in non-blinded and blinded mode: - Training Data (Set A) with non-blinded annotations, including 773 acceptable and 225 unacceptable ECGs; - Test Data (Set B) with blinded annotations, including 500 ECGs. Misplaced electrodes have been manually identified in 74/1498 recordings, publicly available in the list [11].