簡易檢索 / 詳目顯示

回結果列表
研究生： 高金生
Deantana, Christian
論文名稱： 基於深度學習之非侵入性動靜脈瘻管狹窄評估
Non-Invasive Arteriovenous Fistula Stenosis Estimation Using Deep Learning
指導教授： 藍崑展
Lan, Kun-chan
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 資訊工程學系
Department of Computer Science and Information Engineering
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 87
中文關鍵詞： 動靜脈瘻管 、血液透析 、血管通路 、狹窄量化 、深度學習 、音訊分析 、VGGish 、回歸 、雜音聽診
外文關鍵詞： arteriovenous fistula, hemodialysis, vascular access, stenosis quantification, deep learning, audio analysis, VGGish, regression, bruit auscultation
相關次數： 點閱：10  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 動靜脈瘻管（Arteriovenous Fistula, AVF）狹窄是血液透析患者血管通路失效的主要原因之一，然而目前的偵測方法仍存在侵入性高、成本昂貴或主觀性強等問題。先前以聲音為基礎的篩檢系統已證實自動化聽診分析具有可行性，但多數系統將此任務視為二元分類問題，僅能判斷「正常」或「狹窄」，無法提供量化的嚴重程度評估。本論文將動靜脈瘻管監測重新定義為連續性的回歸問題，提出一套非侵入性、僅以音訊為基礎的深度學習模型，直接由聽診器錄音估算狹窄程度（Degree of Stenosis, DOS%）。經心跳分段後的音訊訊號被轉換為對數梅爾頻譜圖（log-mel spectrogram），並以微調後的 VGGish 骨幹網路進行編碼，再透過多層感知器（MLP）回歸頭輸出 DOS% 預測值。為穩定模型在小型臨床資料集上的訓練表現，本研究採用暖啟動（warm-up）訓練、部分凍結網路層、差異化學習率，以及標籤分布平滑（Label Distribution Smoothing）等策略。
    在涵蓋 44 位病患、共 126 筆錄音且以血管攝影為驗證標準的資料集上，本模型於五折交叉驗證中達到平均測試 MAE 為 14.75%，符合依據 KDOQI 準則中 50% 狹窄介入門檻所訂定之臨床目標（MAE ≤ 15%）。頻域分析顯示，632–1000 Hz 頻段的渦流能量為狹窄的主要聲學指標；時域分析則指出收縮期峰值窗口具有最高的判別力，兩者皆與動靜脈瘻管之血流動力學特性相符。上述結果證實，僅憑聲音即可在不依賴影像設備或專業判讀的情況下，量化評估動靜脈瘻管狹窄程度，為未來發展病患自我監測系統奠定基礎；然而，資料集規模、模型泛化能力及實際場域穩健性等限制仍待克服。

    Arteriovenous fistula (AVF) stenosis is a leading cause of vascular access failure in hemodialysis patients, but current detection methods are invasive, costly, or subjective. Sound-based screening has shown promise, but prior systems treat detection as binary classification, giving no measure of severity. This thesis instead frames AVF monitoring as a regression problem, proposing a non-invasive, audio-only pipeline that estimates the Degree of Stenosis (DOS%) from stethoscope recordings. Heartbeat-segmented audio is converted into log-mel spectrograms and encoded by a fine-tuned VGGish backbone, with an MLP head regressing the DOS% output. Warm-up training, partial layer freezing, differential learning rates, and label distribution smoothing are used to stabilize training on the small clinical dataset.
    On 126 recordings from 44 patients, validated against angiography, the model achieved a mean test MAE of 14.75% across five folds, meeting the clinically motivated ≤15% target derived from the KDOQI 50% stenosis threshold. Frequency-domain analysis identifies the 632–1000 Hz band as the primary acoustic marker of stenosis, while time-domain analysis shows the peak-systolic window is most discriminative — both consistent with AVF hemodynamics. These results show that audio-based AVF stenosis quantification is feasible without imaging or specialist expertise, supporting future patient self-monitoring, though dataset size and real-world robustness remain limitations.

    摘要 1 ABSTRACT 2 CONTENTS 3 LIST OF FIGURES 6 LIST OF TABLES 8 1. INTRODUCTION 10 1.1 What is an Arteriovenous Fistula (AVF)? 10 1.2 The Problem with Current Approaches in Detecting AVF Dysfunction 11 1.3 Previous Approaches Using Sound to Detect AVF Dysfunction 12 1.4 Thesis Contributions 13 2. Literature Review 16 2.1 Signal Processing-Based Approaches 16 2.2 Deep Learning-Based Approaches 17 3. Methodology 20 3.1 Architecture 20 3.2 Audio Segmentation 22 3.2.1 The Cardiac Cycle: Systole and Diastole 22 3.2.2 Motivation for Audio Segmentation 23 3.2.3 Segmentation Pipeline 24 3.2.4 Noise Removal 26 3.3 Data Augmentation 27 3.4 Mel-Spectrogram Extraction 28 3.5 VGGish 29 3.6 PPG2ABP Model 30 3.7 Demographic and Physiological Features 31 3.7.1 Age and Heart Rate 31 3.7.2 Blood Pressure 32 3.8 Audio-only / Fusion 32 3.9 MLP Head 34 3.10 Training Strategy 35 3.10.1 VGGish Fine-Tuning Strategy 35 3.10.2 Label Distribution Smoothing (LDS) 36 3.10.3 Stratified Group K-Fold Cross-Validation 38 4. Experiments 40 4.1 Dataset 40 4.1.1 Recording Device and Specifications 40 4.1.2 Data Collection Procedure 40 4.1.3 Patient Population 41 4.1.4 Clinical Labels 42 4.2 Experiment Environment 43 4.3 Experiment Results 44 4.3.1 Performance by Patient Subgroup 53 4.3.2 Association of Demographic and Physiological Variables with Stenosis Severity 58 4.4 Ablation Studies 60 4.4.1 Audio-Only vs. Fusion 60 4.4.2 Warm-Up Strategy 62 4.4.3 Layer Freezing 63 4.4.4 Learning Rate 64 4.4.5 Different Input 65 4.4.6 Different Model 66 4.4.7 Single Heartbeat vs. Multi-Beat Window 67 4.4.8 Systolic Window vs. Full Segment 70 5. Discussion 73 5.1 Frequency-Domain Analysis 73 5.1.1 Frequency Band Masking Method 73 5.1.2 Frequency Analysis Result 74 5.1.3 Clinical Interpretation 74 5.2 Time-Domain Analysis 75 5.2.1 Sliding Window Masking Method 75 5.2.2 Time Analysis Result 75 5.2.3 Clinical Interpretation 77 6. Conclusion 79 7. Limitations and Future Work 80 7.1 Limitations 80 7.2 Future Work 81 REFERENCE 83

    1. Mark, P.B., et al., Global, regional, and national burden of chronic kidney disease in adults, 1990–2023, and its attributable risk factors: a systematic analysis for the Global Burden of Disease Study 2023. The Lancet, 2025. 406(10518): p. 2461–2482.
    2. Lok, C.E., et al., KDOQI clinical practice guideline for vascular access: 2019 update. American Journal of Kidney Diseases, 2020. 75(4): p. S1–S164.
    3. Al-Jaishi, A.A., et al., Complications of the arteriovenous fistula: a systematic review. Journal of the American Society of Nephrology, 2017. 28(6): p. 1839–1850.
    4. MacRae, J.M., et al., Arteriovenous access failure, stenosis, and thrombosis. Canadian journal of kidney health and disease, 2016. 3: p. 2054358116669126.
    5. Segal, M. and E. Qaja, Types of arteriovenous fistulas, in StatPearls [Internet]. 2022, StatPearls Publishing.
    6. Al-Jaishi, A.A., et al., Patency rates of the arteriovenous fistula for hemodialysis: a systematic review and meta-analysis. American Journal of Kidney Diseases, 2014. 63(3): p. 464–478.
    7. Froger, C.L., et al., Stenosis detection with MR angiography and digital subtraction angiography in dysfunctional hemodialysis access fistulas and grafts. Radiology, 2005. 234(1): p. 284–291.
    8. Malik, J., et al., The role of Doppler ultrasonography in vascular access surveillance—controversies continue. The journal of vascular access, 2021. 22(1_suppl): p. 63–70.
    9. Abreo, K., B.M. Amin, and A.P. Abreo, Physical examination of the hemodialysis arteriovenous fistula to detect early dysfunction. The Journal of Vascular Access, 2019. 20(1): p. 7–11.
    10. Gram, M., et al. Stenosis detection algorithm for screening of arteriovenous fistulae. in 15th Nordic-Baltic Conference on Biomedical Engineering and Medical Physics (NBC 2011) 14-17 June 2011, Aalborg, Denmark. 2011. Springer.
    11. Vesquez, P.O., M.M. Marco, and B. Mandersson. Arteriovenous fistula stenosis detection using wavelets and support vector machines. in 2009 Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2009. IEEE.
    12. Sung, P.-H., et al., Hemodialysis vascular access stenosis detection using auditory spectro-temporal features of phonoangiography. Medical & biological engineering & computing, 2015. 53(5): p. 393–403.
    13. Grochowina, M., L. Leniowska, and A. Gala-Błądzińska, The prototype device for non-invasive diagnosis of arteriovenous fistula condition using machine learning methods. Scientific Reports, 2020. 10(1): p. 16387.
    14. Yeih, D.-F., et al. Physiology-based diagnosis algorithm for arteriovenous fistula stenosis detection. in 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2014. IEEE.
    15. Ota, K., et al., Evaluation of hemodialysis arteriovenous bruit by deep learning. Sensors, 2020. 20(17): p. 4852.
    16. Park, J.H., et al., Feasibility of deep learning-based analysis of auscultation for screening significant stenosis of native arteriovenous fistula for hemodialysis requiring angioplasty. Korean journal of radiology, 2022. 23(10): p. 949.
    17. Song, W.T., et al., An effective AI model for automatically detecting arteriovenous fistula stenosis. Scientific Reports, 2023. 13(1): p. 17659.
    18. Zhou, G., et al., Deep learning analysis of blood flow sounds to detect arteriovenous fistula stenosis, NPJ Digit. 2023, Med.
    19. Zhang, J., et al., Classification of arteriovenous fistula sounds using a convolutional block attention module and long short-term memory neural network. Frontiers in Physiology, 2024. 15: p. 1397317.
    20. Chung, T.-L., et al., Prediction of arteriovenous access dysfunction by Mel spectrogram-based deep learning model. International Journal of Medical Sciences, 2024. 21(12): p. 2252.
    21. Poloni, S., et al., The potential of sound analysis to reveal hemodynamic conditions of arteriovenous fistulae for hemodialysis. Annals of Biomedical Engineering, 2025. 53(1): p. 230–240.
    22. Malindretos, P., et al., Computer assisted sound analysis of arteriovenous fistula in hemodialysis patients. 2014, SAGE Publications Sage UK: London, England.
    23. Latif, S., et al., Phonocardiographic sensing using deep learning for abnormal heartbeat detection. IEEE Sensors Journal, 2018. 18(22): p. 9393–9400.
    24. Althaph, B. and N.P. Challa, Explainable attention-based deep learning for classification and interpretation of heart murmurs using phonocardiograms. Scientific Reports, 2025. 15(1): p. 37991.
    25. Sondermann, M., et al. Comparative Analysis of CNN and Transformer Architectures with Heart Cycle Normalization for Automated Phonocardiogram Classification. in 2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). 2025. IEEE.
    26. Munguía, M.M. and B. Mandersson. Analysis of the vascular sounds of the arteriovenous fistula's anastomosis. in 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society. 2011. IEEE.
    27. Chang, H.-Y., et al., Digitized Acoustic Analysis for Monitoring Hemodialysis Access Dysfunction: Insights from Vascular Imaging and Post-Angioplasty Data. Journal of Clinical Medicine, 2026. 15(2): p. 662.
    28. Hershey, S., et al. CNN architectures for large-scale audio classification. in 2017 ieee international conference on acoustics, speech and signal processing (icassp). 2017. IEEE.
    29. Gemmeke, J.F., et al. Audio set: An ontology and human-labeled dataset for audio events. in 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). 2017. IEEE.
    30. Ibtehaz, N., et al., PPG2ABP: Translating photoplethysmogram (PPG) signals to arterial blood pressure (ABP) waveforms. Bioengineering, 2022. 9(11): p. 692.
    31. Yap, Y., et al., Association of early failure of arteriovenous fistula with mortality in hemodialysis patients. Sci Rep 2021; 11 (1): 5699.
    32. Castelli, R., et al., Aging of the arterial system. International journal of molecular sciences, 2023. 24(8): p. 6910.
    33. Gao, M. and J. Wang, [Retracted] Risk Factors of Arteriovenous Fistula Stenosis of Patients with Maintenance Hemodialysis. Evidence‐based Complementary and Alternative Medicine, 2022. 2022(1): p. 2968122.
    34. Sundin, J., et al., Turbulent intensity of blood flow in the healthy aorta increases with dobutamine stress and is related to cardiac output. Frontiers in physiology, 2022. 13: p. 869701.
    35. Ha, H., et al., Age-related vascular changes affect turbulence in aortic blood flow. Frontiers in physiology, 2018. 9: p. 36.
    36. Huang, Y., et al., Impact of Arterial Stiffness on In-Stent Restenosis in the Era of Drug-Eluting Stents. Reviews in Cardiovascular Medicine, 2025. 26(6): p. 23847.
    37. Yosinski, J., et al., How transferable are features in deep neural networks? Advances in neural information processing systems, 2014. 27.
    38. Howard, J. and S. Ruder. Universal language model fine-tuning for text classification. in Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018.
    39. Yang, Y., et al. Delving into deep imbalanced regression. in International conference on machine learning. 2021. PMLR.
    40. Loh, W.Y., Classification and regression trees. Wiley interdisciplinary reviews: data mining and knowledge discovery, 2011. 1(1): p. 14–23.
    41. Roh, K.M., A. Awosika, and R.M. Millis, Advances in wearable stethoscope technology: opportunities for the early detection and prevention of cardiovascular diseases. Cureus, 2024. 16(12).
    42. Asif, A., et al., Inflow stenosis in arteriovenous fistulas and grafts: a multicenter, prospective study. Kidney international, 2005. 67(5): p. 1986–1992.
    43. Gong, Y., Y.-A. Chung, and J. Glass, Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778, 2021.
    44. Baevski, A., et al., wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 2020. 33: p. 12449–12460.

    QR CODE