ecg classification from a short single lead using...
TRANSCRIPT
ECG CLASSIFICATION FROM A SHORT SINGLE LEAD USING MACHINE LEARNING &
HAND-CRAFTED FEATURE EXTRACTIONHeather ShumakerECE 539 • Fall 2018
SUMMARYECG ClassificationNormal, Atrial Fibrillation (AF), Other rhythm, Noisy
• Signal processing
• Feature extraction
• Classifier training & testing
• Committee machine
75% average accuracy
DATA
60%
9%
30%
1%
8,528 ECG SAMPLESNormal AF Other Rhythm Noisy
PROPOSED APPROACH
FINAL APPROACH
Committee machine
1. Normal vs Abnormal classification
2. Abnormality classification
3. Combine results
FEATURE EXTRACTION
Atrial Fibrillation
Other Rhythm
Noisy
MATLAB Classification Learner
RESULTS
1,000 samples used for testing
Accuracy Normal vs. Abnormal (%) AF vs Other vs Noisy (%) Overall
Validation 82.2 80.2
Testing 80.9 61 74.6
DISCUSSIONRoom for improvement
• Uneven class distribution• Better solution for combining classifier outputs• More ECG features • Train neural network on features• Noisy data & inconsistent labeling
Lessons learned
• TensorFlow on laptop for predictions• Python-MATLAB interaction• MATLAB classification learner
QUESTIONS?