Evaluation of an algorithm based on single-condition decision rules for binary classification of 12-lead ambulatory ECG recording quality

A new algorithm for classifying ECG recording quality based on the detection of commonly observed ECG contaminants which often render the ECG unusable for diagnostic purposes was evaluated. Contaminants (baseline drift, flat line, QRS-artefact, spurious spikes, amplitude stepwise changes, noise) were detected on individual leads from joint time-frequency analysis and QRS amplitude. Classification was based on cascaded single-condition decision rules (SCDR) that tested levels of contaminants against classification thresholds. A supervised learning classifier (SLC) was implemented for comparison. The SCDR and SLC algorithms were trained on an annotated database (Set A, PhysioNet Challenge 2011) of 'acceptable' versus 'unacceptable' quality recordings using the 'leaveMout' approach with repeated random partitioning and cross-validation. Two training approaches were considered: (i) balanced, in which training records had equal numbers of 'acceptable' and 'unacceptable' recordings, (ii) unbalanced, in which the ratio of 'acceptable' to 'unacceptable' recordings from Set A was preserved. For each training approach, thresholds were calculated, and classification accuracy of the algorithm compared to other rule based algorithms and the SLC using a database for which classifications were unknown (Set B PhysioNet Challenge 2011). The SCDR algorithm achieved the highest accuracy (91.40%) compared to the SLC (90.40%) in spite of its simple logic. It also offers the advantage that it facilitates reporting of meaningful causes of poor signal quality to users.