Enhancing Arrhythmia Classification: A Comparative Study of Filters and Deep Learning Models on ECG Data

Arrhythmia classification using electrocardiogram (ECG) data is crucial in the treatment of cardiovascular disease. However, ECG data often includes noise, which may conceal essential facts, making categorization difficult. This paper compares a notch filter with Gaussian noise, a Finite Impulse Response (FIR) filter, and an Infinite Impulse Response (IIR) filter using a 1D convolutional neural network (1DCNN) and DenseNet121. We obtained the ECG signals for this investigation from the MIT-BIH arrhythmia database. The preprocessing pipeline incorporates Gaussian noise into the notch filter at various frequency bands. The proposed IIR filter addresses baseline wander, power-line interference, and muscle artifacts, whereas FIR filters minimize high-frequency noise and baseline wander. Then, we train the 1DCNN and DenseNet-121 models on the filtered ECG data. These models classify arrhythmias as normal (N), supraventricular (S), ventricular (V), fusion (F), or unknown (Q). Accuracy, precision, recall, and F1 score are standard metrics used to assess model-filter combinations. Each filtering strategy increases classification performance, and 1DCNN and DenseNet-121 outperform in noisy ECG data. The IIR filter with the DenseNet-121 method achieved the best accuracy of 9 9. 9 7 %, while the Gaussian noise with the notch filter using the 1DCNN method got 99%, contributing to diagnostic capabilities. These discoveries might help detect arrhythmias and improve patient outcomes in cardiovascular diagnostics.