Classification of Arrhythmias Using a Cnn-patch Time Series Transformer via ECG Signals

Ramadhan, Humam Maulana Tsubasanofa (2026) Classification of Arrhythmias Using a Cnn-patch Time Series Transformer via ECG Signals. Undergraduate thesis, UPN Veteran Jawa Timur.

This is the latest version of this item.

	Text (Cover) 20081010084-cover.pdf Download (2MB)
	Text (Bab 1) 20081010084-bab1.pdf Download (129kB)
	Text (Bab 2) 20081010084-bab2.pdf Restricted to Repository staff only until 22 July 2029. Download (1MB)
	Text (Bab 3) 20081010084-bab3.pdf Restricted to Repository staff only until 22 July 2029. Download (1MB)
	Text (Bab 4) 20081010084-bab4.pdf Restricted to Repository staff only until 22 July 2029. Download (1MB)
	Text (Bab 5) 20081010084-bab5.pdf Download (87kB)
	Text (Daftar pustaka) 20081010084-daftarpustaka.pdf Download (125kB)

Abstract

Arrhythmia, or cardiac rhythm disorder, is one of the cardiovascular conditions that can lead to stroke, heart failure, and sudden cardiac death. Traditional arrhythmia identification is performed through manual observation of electrocardiogram (ECG) signal recordings, encompassing analysis of rhythm, duration, orientation, and waveform morphology. This process is time-consuming and prone to interpretation errors, prompting the development of various machine learning models to assist clinicians in detecting and classifying ECG waveform patterns. Deep Learning techniques such as CNN and RNN have demonstrated potential in automating ECG analysis, yet each carries notable limitations: CNNs often struggle to capture global temporal dependencies, while RNNs are susceptible to vanishing gradient problems on long sequences. This study explores the application of a hybrid end-to-end CNN-PatchTST architecture, where the CNN handles local morphological feature extraction from ECG signals before being processed by PatchTST, which leverages a patching mechanism to overcome the quadratic complexity of self-attention and efficiently capture long-range temporal dependencies. The model was evaluated on five arrhythmia classes — Atrial Fibrillation (AF), First-degree Atrioventricular Block (IAVB), Normal Sinus Rhythm (NSR), Sinus Bradycardia (SB), and Sinus Tachycardia (STach) — using multiple public ECG datasets with varying look-back window configurations (L=336 and L=512) and patch configurations (patch_length=16/stride=8 and patch_length=32/stride=16). Experimental results show that the best configuration, L=336 with patch_length=32 and stride=16 at a 70:20:10 split ratio, achieves an accuracy of 88.8%, macro sensitivity of 84.3%, macro specificity of 96.3%, and macro F1-score of 82.5%. Look-back window L=336 consistently outperforms L=512 in both classification performance and training efficiency, reducing training time by up to 77% under the same patch configuration. Per-class analysis reveals significant performance variation, with NSR and AF achieving the highest F1-scores of 93% and 91% respectively, while IAVB and SB obtain the lowest at 70% and 73%, reflecting the inherent challenge of distinguishing classes with high morphological similarity. Keywords: Arrhythmia, Electrocardiogram, Convolutional Neural Network, PatchTST, Transformer

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Junaidi, Achmad	NIDN0710117803	achmadjunaidi.if@upnjatim.ac.id
Thesis advisor	Haromainy, Muhammad Muharrom Al	NIDN0701069503	muhammad.muharrom.if@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA76.6 Computer Programming
T Technology > T Technology (General)

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Humam Maulana Tsubasanofa Ramadhan

Date Deposited:

22 Jul 2026 06:51

Last Modified:

22 Jul 2026 07:52

URI:

https://repository.upnjatim.ac.id/id/eprint/57720