ECG Arrhythmia Classification

MIT-BIH Dataset · Feature Engineering · Data Balancing · Augmentation · SVM Classifier

This project implements a complete classification pipeline for ECG heartbeat signals using the MIT-BIH Arrhythmia dataset.
The goal is to classify heartbeats into five arrhythmia types using rhythmic features, data balancing, augmentation techniques, and an SVM classifier with an RBF kernel.

---

Dataset Overview

The dataset contains ECG beats labeled into 5 classes:

• N — Normal
• S — Fusion of paced and normal beats
• V — Premature Ventricular Contraction
• F — Atrial premature contraction
• Q — Fusion of ventricular and normal beats

Each row contains 187 ECG samples + 1 class label.

Dataset size:

• Training set: 87,553 samples
• Test set: 21,891 samples

---

Dataset Exploration

Initial operations included:

1. Loading the CSV datasets using Pandas
2. Inspecting structure and label distribution
3. Visualizing ECG waveforms
4. Identifying severe class imbalance (class N dominates)

Waveform visualization was performed using Matplotlib by sampling signals at fixed intervals.

---

Subsampling

Due to computational constraints, 10% of the original training and testing sets were extracted using:

train_test_split(..., stratify=labels)

Resulting shapes:

• Train: 8755 × 188
• Test: 2189 × 188

---

Preprocessing

Normalization

Signals were normalized with MinMaxScaler (-1, 1):

• fit_transform() → training set
• transform() → test set

Feature/Label Separation

The last column was split from the ECG samples.

---

Feature Extraction

Feature extraction was performed through the compute_feature_vector() function, producing rhythmic and statistical descriptors.

Extracted features:

Statistical features

• Mean
• Standard deviation

Zero-Crossing Rate (ZCR)

• Frame-wise ZCR
• Mean of ZCR
• Standard deviation of ZCR

Short-Time Fourier Transform (STFT)

• Magnitude spectrum

Spectral Flux

• Spectral flux using librosa.onset.onset_strength
• Mean and standard deviation

All features (statistics + frame features) are concatenated into a single vector.

Final shapes:

• Train feature matrix: (8755, 30)
• Test feature matrix: (2189, 30)

---

Dataset Balancing

To remove bias toward class N, a perfectly balanced dataset was created:

• Train: 641 samples per class → 3205 samples
• Test: 162 samples per class → 810 samples

This was done via random sampling for each class.

---

Data Augmentation

Two augmentation techniques were implemented:

1. Stretch

Non-linear resampling and padding/truncation to 187 samples.

2. Amplify

Amplitude scaling with a factor α ∈ [-0.5, 0.5].

Augmentation strategy

The perform() method applies stretch and/or amplify with 50% probability.

For each class:

• 100 augmented samples were generated
• Added to the balanced dataset

Final augmented training size:

• 3705 samples

Normalization was re-applied afterward.

---

SVM Classifier

Classifier setup:

• Kernel: RBF
• C: 10 (initial run)

Performance (balanced dataset):

• Test accuracy: 67.53%
• Train accuracy: 65.75%
  → No overfitting.

---

Evaluation Metrics

Classification Report (balanced test set — 810 samples)

Class	Precision	Recall	F1-score
N	0.59	0.69	0.64
S	0.84	0.67	0.75
V	0.80	0.71	0.75
F	0.79	0.89	0.84
Q	0.89	0.91	0.90

• Overall accuracy: 78%
• Macro F1: 0.78

Compared to the original imbalanced dataset, minority class performance improved dramatically.

---

Hyperparameter Tuning

Several C values were tested:

C	Accuracy	Macro F1	Macro Recall	False Negatives
0.1	67.53%	0.6666	0.6753	263
1	72.72%	0.7252	0.7272	221
100	79.26%	0.7926	0.7926	168
1000	80.00%	0.8003	0.8000	162
10000	77.28%	0.7732	0.7728	184
100000	77.04%	0.7722	0.7704	186

Best configuration:

• Kernel: RBF
• C: 1000
• Gamma: auto
• N: 64
• H: 8
• Accuracy: 80.49%

---

Final Notes and Future Work

Key takeaways

• Balancing the dataset is crucial.
• Rhythmic features significantly help distinguish arrhythmias.
• Augmentation improves generalization.
• SVM with RBF delivers robust results (~80%).

Potential improvements

• 1D CNNs for automatic feature extraction
• Additional augmentations (noise, shifts, warping)
• PCA or feature selection
• Class-weighted SVM or focal loss