Scalable Noise Mining in Long-Term Electrocardiographic Time-Series to Predict Death Following Heart Attacks Chih-Chun Chia, Zeeshan Syed University of Michigan Presenter: Sen Jiao Mar. 19, 2015
Objective Improve an algorithm for mining useful information in electrocardiogram (ECG) to identify patients at an increased risk of death following heart attacks.
Introduction Heart disease: 34% of all deaths each year in the U.S, 1 death in every 38 seconds Reducing mortality: inability to match patients to treatments that are most appropriate for individual risk. Implantable cardioverter defibrillator (ICD) Current decision-making methods fail to prescribe ICD to the majority.
Introduction Biomarkers: to estimate patient risk and to match patients to treatments. Blood-based measurements, medical imaging: limited to available information. ECG may contain subtle but useful information, commonly perceived as noise. Morphologic variability (MV) in ECG: indicator of heart function.
Background Electrocardiogram (ECG)
Background Pathophysiology
Background Morphological Variability (MV)
Background Dynamic Time-Warping (DTW)
Background Dynamic Time-Warping (DTW)
Background Constrained DTW: prevent biologically implausible alignments – Boundary Conditions – Continuity – Monotonicity
Background Constrained DTW
Background Power Spectral Density – Lomb-Scargle periodogram
Overall Flow Chart O(pn 2 )
Adaptive Down-sampling (ADAP) PAA, FastDTW,ADAP
DTW with ADAP
Evaluation 4-day continuous ECG data recorded Patients follow-up for 90 days for cardiovascular death Evaluate MV, MV measured with PAA, MV measured with ADAP Measure areas under the receiver operating characteristic curves (AUROCs)
Results
Conclusion ADAP substantially reduces runtime while providing similar performance to the basic MV algorithm that is not optimized for large volumes of data. The use of ADAP leads to more accurate performance than downsampling through the commonly used approach of PAA.
Q&A Thank you!