ECG anomaly class identification using LSTM and error profile modeling.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning, particularly Long Short Term Memory (LSTM) networks, shows promise in automatic cardiac event diagnosis using electrocardiogram (ECG) data.
• Previous work utilized LSTM prediction errors on normal ECGs for anomaly detection.
• Real-time diagnostic applications require accurate identification of specific cardiac anomaly types.

Purpose of the Study:

• To extend an existing anomaly detection algorithm by introducing a second-stage predictor for classifying specific cardiac anomaly types.
• To evaluate the effectiveness of different machine learning models as the second-stage predictor.
• To propose and validate a featurization scheme for LSTM prediction errors to improve classification performance.

Main Methods:

• Implemented a two-stage deep learning approach: an initial LSTM network for anomaly detection and a second-stage predictor for classification.
• Employed multiple second-stage models including multilayer perceptron (MLP), support vector machine (SVM), and logistic regression.
• Developed a featurization scheme summarizing LSTM prediction errors and trained a predictor using these summary features.

Main Results:

• The second-stage predictor successfully identified seven types of cardiac anomalies: Atrial Premature Contraction (APC), Paced Beat (PB), Premature Ventricular Contraction (PVC), Right Bundle Branch Block (RBBB), Ventricular Bigeminy (VB), Ventricular Couplets (VCs), and Ventricular Tachycardia (VT).
• The proposed featurization scheme for LSTM prediction errors proved effective, with the summary features carrying significant predictive information about the ECG anomaly type.
• Performance varied across different second-stage models, indicating the importance of model selection for optimal anomaly class prediction.

Conclusions:

• The enhanced two-stage deep learning model effectively classifies various cardiac anomalies from ECG data.
• Summary features of LSTM prediction errors are valuable for identifying specific ECG anomaly types.
• Accurate performance estimation requires careful consideration of class imbalances and data paucity, necessitating robust background models for multi-class predictors.