DEVELOPMENT OF AN ATRIAL FIBRILLATION DETECTION ALGORITHM FOR ECG COLLECTED WITH TEXTILE GARMENT

BackgroundAtrial fibrillation (AF) is a common arrhythmia related to increased mortality, strokes and heart failure. AF is typically diagnosed using ECG recorded with adhesive gel electrodes, which can irritate the skin in long recording. Garments with textile electrodes are a comfortable alternative for long term ECG monitoring.ObjectiveTo develop and evaluate an algorithm for continuous monitoring and automatic detection of AF using ECG recording garments.MethodsThe proposed AF detection from textile ECG includes 3 steps: (1) R-peak detection using a proprietary algorithm previously developed and validated for textile ECG; (2) Signal quality classification (clean/noisy ECG segments) to exclude expected noise and motion artifacts with a threshold on the ratio [maximum slope in between R-peak regions]/[maximum slope of the corresponding R-peaks]; and (3) Classification of AF and non-AF based on the Shannon Entropy of the processed RR sequence (using an optimized threshold) of the clean ECG segments (with optimized segment length). Initial development and testing of the AF detection algorithm was performed using public databases. The MIT-AFIB public database was used to optimize the segment length and entropy threshold to discriminate between AF and non-AF segments. Final testing was done with ECG recorded from Skiin Undergarment (Myant Inc, Canada) in 84 participants and labeled by a cardiologist in an REB-approved study (PACE cardiology, Canada).ResultsThe optimal parameters for this AF detection method are a segment duration of 55 sec and an entropy threshold of 0.95. The sensitivity and specificity of AF detection for each database used in this study was above 91% as detailed in Table 1.ConclusionTable 1Results of the Atrial Fibrillation detection algorithm for training and testing datasets. One sample here refer to 55 seconds of ECG assigned the given label. Specificity is reported separately in two conditions for completness.DatabaseMIT-AFib (training, gel ECG)MIT-arrhythmia (testing, gel ECG)Long term AFib (testing, gel ECG)PACE study (testing, textile ECG)Number of AF samples459814139267138Number of Normal samples6648948387921355Number of Other arrhythmia samples1014471144084Sensitivity of AF detection95.3%98.5%92.2%91.3%Specificity of AF detection for AF vs Normal only97.5%96.9%99.4%91.3%Specificity of AF detection for AF vs Normal and Other97.3%89.0%94.9%91.1% Open table in a new tab BackgroundAtrial fibrillation (AF) is a common arrhythmia related to increased mortality, strokes and heart failure. AF is typically diagnosed using ECG recorded with adhesive gel electrodes, which can irritate the skin in long recording. Garments with textile electrodes are a comfortable alternative for long term ECG monitoring. Atrial fibrillation (AF) is a common arrhythmia related to increased mortality, strokes and heart failure. AF is typically diagnosed using ECG recorded with adhesive gel electrodes, which can irritate the skin in long recording. Garments with textile electrodes are a comfortable alternative for long term ECG monitoring. ObjectiveTo develop and evaluate an algorithm for continuous monitoring and automatic detection of AF using ECG recording garments. To develop and evaluate an algorithm for continuous monitoring and automatic detection of AF using ECG recording garments. MethodsThe proposed AF detection from textile ECG includes 3 steps: (1) R-peak detection using a proprietary algorithm previously developed and validated for textile ECG; (2) Signal quality classification (clean/noisy ECG segments) to exclude expected noise and motion artifacts with a threshold on the ratio [maximum slope in between R-peak regions]/[maximum slope of the corresponding R-peaks]; and (3) Classification of AF and non-AF based on the Shannon Entropy of the processed RR sequence (using an optimized threshold) of the clean ECG segments (with optimized segment length). Initial development and testing of the AF detection algorithm was performed using public databases. The MIT-AFIB public database was used to optimize the segment length and entropy threshold to discriminate between AF and non-AF segments. Final testing was done with ECG recorded from Skiin Undergarment (Myant Inc, Canada) in 84 participants and labeled by a cardiologist in an REB-approved study (PACE cardiology, Canada). The proposed AF detection from textile ECG includes 3 steps: (1) R-peak detection using a proprietary algorithm previously developed and validated for textile ECG; (2) Signal quality classification (clean/noisy ECG segments) to exclude expected noise and motion artifacts with a threshold on the ratio [maximum slope in between R-peak regions]/[maximum slope of the corresponding R-peaks]; and (3) Classification of AF and non-AF based on the Shannon Entropy of the processed RR sequence (using an optimized threshold) of the clean ECG segments (with optimized segment length). Initial development and testing of the AF detection algorithm was performed using public databases. The MIT-AFIB public database was used to optimize the segment length and entropy threshold to discriminate between AF and non-AF segments. Final testing was done with ECG recorded from Skiin Undergarment (Myant Inc, Canada) in 84 participants and labeled by a cardiologist in an REB-approved study (PACE cardiology, Canada). ResultsThe optimal parameters for this AF detection method are a segment duration of 55 sec and an entropy threshold of 0.95. The sensitivity and specificity of AF detection for each database used in this study was above 91% as detailed in Table 1. The optimal parameters for this AF detection method are a segment duration of 55 sec and an entropy threshold of 0.95. The sensitivity and specificity of AF detection for each database used in this study was above 91% as detailed in Table 1. ConclusionTable 1Results of the Atrial Fibrillation detection algorithm for training and testing datasets. One sample here refer to 55 seconds of ECG assigned the given label. Specificity is reported separately in two conditions for completness.DatabaseMIT-AFib (training, gel ECG)MIT-arrhythmia (testing, gel ECG)Long term AFib (testing, gel ECG)PACE study (testing, textile ECG)Number of AF samples459814139267138Number of Normal samples6648948387921355Number of Other arrhythmia samples1014471144084Sensitivity of AF detection95.3%98.5%92.2%91.3%Specificity of AF detection for AF vs Normal only97.5%96.9%99.4%91.3%Specificity of AF detection for AF vs Normal and Other97.3%89.0%94.9%91.1% Open table in a new tab