Klasifikasi Abnormalitas Paru-Paru Manusia Berbasis Frekuensi Suara Menggunakan YAMNet

Arifin, Jaenal (2025) Klasifikasi Abnormalitas Paru-Paru Manusia Berbasis Frekuensi Suara Menggunakan YAMNet. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7022221021 Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Penyakit paru-paru masih menjadi penyumbang utama morbiditas dan mortalitas global. Auskultasi konvensional menggunakan stetoskop bersifat subjektif dan sangat bergantung pada kompetensi klinis tenaga medis. Kondisi ini membuka celah riset dalam melakukan klasifikasi abnormalitas suara paru-paru manusia. Penelitian ini mengkaji pengembangan sistem klasifikasi suara paru-paru melalui eksplorasi dan integrasi beberapa pendekatan pembelajaran transfer dan representasi sinyal audio. Tiga pendekatan utama dianalisis, meliputi pemanfaatan YAMNet sebagai pre-trained audio feature extractor, penerapan transfer learning berbasis ResNet50 menggunakan representasi mel spectrogram, serta penggabungan YAMNet dengan neural networks dan teknik augmentasi data audio. Sinyal suara paru-paru diproses dalam domain waktu–frekuensi untuk mengekstraksi karakteristik spektral yang relevan secara fisiologis. Strategi fine-tuning pada model ResNet50 diterapkan untuk menyesuaikan representasi fitur pra-latih terhadap karakteristik spesifik suara paru-paru. Hasil evaluasi menunjukkan peningkatan kinerja yang signifikan. Pendekatan transfer learning berbasis ResNet50 dengan pengoptimal adamax menghasilkan performa terbaik dengan akurasi 95,33%, presisi 95,46%, recall 93,33%, F1-score 95,24%, dan AUC mencapai 99,63%. Model yang diusulkan mampu melakukan klasifikasi secara sempurna (AUC 1.00) pada kategori suara rhonchi. Selain itu, penggunaan teknik validasi silang pada arsitektur hibrida mencapai akurasi tertinggi sebesar 97,24%, sementara integrasi dataset kombinasi (ICBHI dan Mendeley) menghasilkan nilai presisi dan akurasi mencapai 100%. Temuan ini menegaskan potensi implementasi model yang diusulkan sebagai bagian dari sistem pendukung keputusan klinis untuk meningkatkan objektivitas dan efisiensi diagnosis penyakit paru-paru di masa depan.
Kata kunci: Klasifikasi abnormalitas, paru-paru manusia, frekuensi suara, YAMNet
===================================================================================================================================
Lung disease remains a major contributor to global morbidity and mortality. Conventional auscultation using a stethoscope is subjective and highly dependent on the clinical competence of medical personnel. This situation creates a research gap in the classification of human lung-sound abnormalities. This study examines the development of a lung sound classification system by exploring and integrating several transfer learning approaches and audio signal representations. Three main approaches are analyzed: the use of YAMNet as a pre-trained audio feature extractor, the application of ResNet50-based transfer learning using mel spectrogram representations, and the combination of YAMNet with neural networks and audio data augmentation techniques. Lung sound signals are processed in the time-frequency domain to extract physiologically relevant spectral characteristics. Lung sound signals are processed in the time-frequency domain to extract physiologically relevant spectral characteristics. A fine-tuning strategy is applied to the ResNet50 model to adapt the pre-trained feature representations to the specific characteristics of lung sounds. The evaluation results showed significant performance improvements. The ResNet50-based transfer learning approach with the Adamax optimizer produced the best performance with an accuracy of 95.33%, a precision of 95.46%, a recall of 93.33%, an F1-score of 95.24%, and an AUC of 99.63%. The proposed model was able to perform perfect classification (AUC 1.00) in the rhonchi sound category. Furthermore, the use of cross-validation techniques in the hybrid architecture achieved the highest accuracy of 97.24%, while the integration of the combined datasets (ICBHI and Mendeley) resulted in precision and accuracy values reaching 100%. These findings confirm the feasibility of implementing the proposed model as part of a clinical decision support system to improve the objectivity and efficiency of lung disease diagnosis.
Keywords : Classification of abnormalities, human lungs, sound frequency, YAMNet

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Klasifikasi abnormalitas, paru-paru manusia, frekuensi suara, YAMNet Classification of abnormalities, human lungs, sound frequency, YAMNet
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5102.9 Signal processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Jaenal Arifin
Date Deposited:	15 Jan 2026 05:21
Last Modified:	15 Jan 2026 05:21
URI:	http://repository.its.ac.id/id/eprint/129640

Actions (login required)

View Item