Klasifikasi Kelainan Jantung pada Citra Elektrokardiogram (EKG) Menggunakan Deep Learning

Putri, Anneu Tsabita (2025) Klasifikasi Kelainan Jantung pada Citra Elektrokardiogram (EKG) Menggunakan Deep Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

This is the latest version of this item.

[thumbnail of 5025211026-Undergraduate_Thesis.pdf] Text
5025211026-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only
Download (9MB) | Request a copy

Abstract

Interpretasi Elektrokardiogram (EKG) yang akurat sangat krusial dalam penanganan pasien di Instalasi Gawat Darurat, namun variabilitas antar dokter dan probabilitas human error dapat membahayakan nyawa pasien, dengan penelitian menunjukkan lebih dari 30% tenaga medis menghasilkan interpretasi yang kurang akurat. Penelitian ini mengembangkan sistem klasifikasi EKG berbasis deep learning menggunakan model Convolutional Neural Network (CNN) yang lightweight dan efisien untuk implementasi mobile, dengan membandingkan 5 arsitektur terbaru yaitu ConvNeXt V2, EfficientNet V2, RegNetY, MobileNet V3, dan MixNet. Pemilihan arsitektur CNN mempertimbangkan efisiensi (lightweight) serta arsitektur yang telah diperbarui atau dikembangkan berdasarkan temuan dan praktik terkini, untuk mengadopsi peningkatan performa dan efisiensi dari model-model sebelumnya. Metodologi penelitian meliputi tahapan pengumpulan dataset publik berisi 938 citra EKG dari empat kelas kondisi jantung, praproses citra dengan resizing ke 224×224 piksel dan normalisasi berdasarkan statistik ImageNet, serta empat skenario uji coba meliputi hyperparameter tuning tanpa pretrained weights, transfer learning dengan pretrained weights, progressive fine-tuning, dan pengujian robustness dengan data augmentasi. Hasil penelitian menunjukkan bahwa EfficientNet V2 dengan transfer learning mencapai performa terbaik dengan akurasi 98,94%, precision 98,78%, recall 98,78%, dan F1-Score 98,78%, dimana kelas myocardial_infarction dan normal berhasil diklasifikasikan dengan sempurna (100% akurasi) sementara kesalahan utama terjadi pada pembedaan antara kelas abnormal_heartbeat dan history_of_MI yang memiliki karakteristik EKG serupa, serta model menunjukkan robustness yang baik terhadap berbagai jenis noise visual dengan tetap mempertahankan performa meskipun PSNR turun hingga 6,25 dB.
==================================================================================================================================
Accurate interpretation of electrocardiograms (ECG) is crucial in emergency department settings, as variability among physicians and the potential for human error can pose serious risks to patients’ lives. Studies have shown that over 30% of medical personnel produce inaccurate ECG interpretations. This study develops an ECG classification system based on deep learning using lightweight and efficient Convolutional Neural Network (CNN) models suitable for mobile implementation. Five recent architectures were compared: ConvNeXt V2, EfficientNet V2, RegNetY, MobileNet V3, and MixNet. The selection of CNN architectures considers efficiency (lightweight design) as well as architectures that have been updated or developed based on recent findings and best practices, in order to adopt performance and efficiency improvements over previous models. The methodology includes collecting a public dataset consisting of 938 ECG images from four heart condition classes, preprocessing by resizing to 224×224 pixels and normalization using ImageNet statistics, and evaluating four experimental scenarios: hyperparameter tuning without pretrained weights, transfer learning with pretrained weights, progressive fine-tuning, and robustness testing using data augmentation. Results show that EfficientNet V2 with transfer learning achieved the best performance, with 98,94% accuracy, 98,78% precision, 98,78% recall, and 98,78% F1-Score. The myocardial infarction and normal classes were perfectly classified (100% accuracy), while the main misclassifications occurred between the abnormal_heartbeat and history_of_MI classes, which share similar ECG characteristics. Additionally, the model demonstrated strong robustness to various types of visual noise, maintaining high performance even when PSNR dropped to 6,25 dB.

Item Type: Thesis (Other)
Uncontrolled Keywords: CNN, Deep Learning, EfficientNet V2, Elektrokardiogram, Kelainan Jantung, CNN, Deep Learning, EfficientNet V2, Electrocardiogram, Heart Abnormalities
Subjects: R Medicine > RC Internal medicine > RC683.5.E5 Electrocardiography
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User: ANNEU TSABITA PUTRI
Date Deposited: 08 Aug 2025 04:07
Last Modified: 08 Aug 2025 04:07
URI: http://repository.its.ac.id/id/eprint/123970

Available Versions of this Item

Actions (login required)

View Item View Item