Speech-Based Detection and Severity Estimation of Chronic Kidney Disease = 음성을 활용한 만성 신장 질환의 진단 및 중증도 평가

국문 초록 (Abstract)

만성콩팥병(Chronic Kidney Disease, CKD)은 신장 기능이 점진적으로 약화되는 질환으로, 전 세계적으로 높은 유병률과 임상적 부담을 지니고 있음에도 불구하고 조기 진단이 어렵다. 현재까지의 진단은 대부분 혈액 및 소변 검사를 기반으로 하며, 이는 침습적이고 비용 및 접근성 측면에서 제약이 있다. CKD는 호흡계 및 발성기관에 생리적 영향을 미치므로, 음성은 잠재적은 비침습적 생체지표로 활용될 수 있다. 그러나 기존까지 CKD에 특화된 음성 코퍼스나, CKD 음성의 포괄적인 음향 분석, 자동 진단 시스템은 존재하지 않았다.
본 학위논문은 이러한 공백을 해소하기 위해, CKD의 음성 기반 진단 및 중증도 예측을 위한 종합적인 분석 파이프라인을 제안한다. 본 연구는 (1) 생리적 근거에 기반한 비침습적 CKD 바이오마커로서 음성의 활용 가능성 탐색, (2) CKD 특화 대규모 음성 코퍼스 구축, (3) 호흡, 발성, 유창성 영역에 대한 정밀 음향 분석 수행, (4) 머신러닝 및 딥러닝 기반의 자동 진단 모델 개발과 해석을 포함한다.
통계적 분석을 통해, CKD 화자의 음성에서 호흡 지지력 저하, 과긴장 발성, 발성 불안정성 등 주요 음향적 특징을 확인하였으며, 반면 CKD 중증도 간에는 일관된 차이가 나타나지 않음을 입증하였다. 임상적으로 검증된 특징들을 입력으로 활용한 머신러닝 기반 모델은 CKD 탐지에서 0.9308의 macro F1-score를 달성하였고, SHAP 해석을 통해 발성 안정성과 발화 유창성 관련 특징들이 모델 예측에 핵심적으로 기여함을 확인하였다. 더불어, 스펙트로그램과 성문 정보를 통합한 Transformer 기반 딥러닝 모델을 새롭게 제안하였으며, 이 모델은 CKD 탐지에서 0.9207의 macro F1-score를 달성하며 모든 제거 (ablation) 구성보다 우수한 성능을 보였다. 모델 해석을 위해 attention weight 시각화, gradient 기반 saliency 분석, Grad-CAM++ 등을 활용하여, 예측에 기여하는 생리학적으로 타당한 음성 패턴을 규명하였다.
결론적으로, 본 논문은 음성을 활용한 비침습적이고 확장 가능한 CKD 자동 진단의 새로운 가능성을 제시한다. 본 연구는 신장 질환 분야에서 디지털 바이오마커로서의 음성의 임상적 활용 가능성을 입증하며, 조기 진단의 효율성과 접근성을 높이는 데 기여할 수 있다.

다국어 초록 (Multilingual Abstract)

Chronic Kidney Disease (CKD) is a progressive condition characterized by a decline in kidney function. Despite its global prevalence and clinical burden, early diagnosis remains challenging due to the reliance on invasive and resource-intensive assessments such as blood and urine tests. Given that CKD affects respiratory and phonatory systems, speech production may reflect underlying physiological impairments. However, prior to this work, no dedicated CKD speech corpus, comprehensive acoustic analysis, or automated diagnostic framework based on speech data had been established.
To address these gaps, this dissertation presents a comprehensive analytical pipeline for speech-based CKD detection and severity estimation. The pipeline includes: (1) motivating the use of speech as a physiologically grounded, non-invasive biomarker for CKD, (2) constructing a large-scale CKD speech corpus containing sustained vowels, sentences, and paragraph readings, (3) conducting detailed acoustic analyses across respiratory, phonatory, and timing/fluency domains, and (4) developing and interpreting machine learning- and deep learning-based diagnostic models.
Through statistical analyses, we identified key acoustic markers of CKD-such as impaired respiratory support, pressed phonation, and phonatory instability-while confirming the limited utility of speech features for disease staging. Machine learning models trained on statistically derived features achieved a macro F1-score of 0.9308 for CKD detection, and SHAP-based interpretation highlighted the diagnostic relevance of features associated with phonatory stability and timing. Additionally, we proposed a novel deep learning framework integrating spectrogram and glottal features via Transformer-based attention fusion. This model achieved a macro F1-score of 0.9207 and outperformed ablated configurations. Interpretability analyses, including attention weights, gradient-based saliency, and Grad-CAM++, revealed physiologically grounded patterns that drove model predictions.
In conclusion, this dissertation introduces a new paradigm for non-invasive, scalable CKD diagnosis using speech. The proposed methods lay a foundation for speech-based digital biomarkers in renal healthcare, with the potential to enhance early detection, accessibility, and clinical decision-making.

목차 (Table of Contents)

• Abstract i
• Chapter 1 Introduction １
• 1.1 Motivation １
• 1.1.1 Clinical Significance of Chronic Kidney Disease １
• 1.1.2 Diagnostic Limitations in CKD ２
• 1.1.3 Voice as a Diagnostic Medium ３
• 1.2 Research Objectives and Dissertation Contributions ４
• 1.3 Dissertation Structure ７
• Chapter 2 Related Studies ９
• 2.1 An Overview of Speech-Based Disease Diagnosis ９
• 2.1.1 Machine Learning Approach １１
• 2.1.2 Deep Learning Approach １３
• 2.2 Acoustic Characteristics of CKD Speech １５
• 2.3 Limitations of Existing CKD Studies and Research Needs １８
• Chapter 3 Constructing Speech Corpora of CKD ２１
• 3.1 Objectives ２１
• 3.2 Speech Data Collection Protocol ２２
• 3.2.1 Participants and Metadata ２２
• 3.2.2 Speech Stimuli ２５
• 3.2.3 Recording Procedure ２６
• 3.3 Summary ２７
• Chapter 4 Acoustic Analysis of CKD Speech ２９
• 4.1 Objectives ２９
• 4.2 Feature Design and Categorization ２９
• 4.2.1 Respiration-Related Features ３２
• 4.2.2 Phonation-Related Features ３３
• 4.2.3 Timing and Fluency Features ３７
• 4.3 Statistical Analyses ３８
• 4.3.1 Dataset Partitioning for Acoustic Analysis ３８
• 4.3.2 Methods ３８
• 4.3.3 Results ４２
• 4.4 Discussion ４７
• 4.4.1 Characteristics of CKD Speech ４８
• 4.4.2 Acoustic Features According to CKD Severity ５１
• 4.5 Summary ５３
• Chapter 5 Machine Learning-Based CKD Detection and Severity Estimation ５５
• 5.1 Objectives ５５
• 5.2 Methods ５５
• 5.2.1 Feature Selection ５５
• 5.2.2 Classification Models for CKD Detection ５９
• 5.2.3 Ordinal Regression Models for CKD Severity Estimation ６０
• 5.3 Experimental Design ６１
• 5.3.1 Data Partitioning and Cross-Validation ６１
• 5.3.2 Feature Selection Strategies ６２
• 5.3.3 Model Training and Hyperparameter Tuning ６３
• 5.3.4 Evaluation Metrics ６４
• 5.4 SHAP-Based Model Interpretability ６５
• 5.5 Results ６７
• 5.5.1 CKD Detection Task ６７
• 5.5.2 CKD Severity Estimation Task ７１
• 5.5.3 SHAP-Analysis Results ７４
• 5.6 Discussion ７８
• 5.6.1 Acoustic Feature-Based CKD Detection ７８
• 5.6.2 Acoustic Feature-Based CKD Severity Estimation ７９
• 5.6.3 Model Interpretation via SHAP Analysis ８１
• 5.7 Summary ８３
• Chapter 6 Hybrid Acoustic Modeling for Deep CKD Detection and Severity Estimation ８６
• 6.1 Objectives ８６
• 6.2 Model Architecture ８８
• 6.2.1 Feature Extraction ８８
• 6.2.2 Transformer-Based Feature Fusion ９０
• 6.2.3 Output Heads for CKD Detection and Severity Estimation ９１
• 6.3 Experimental Design ９３
• 6.3.1 Data ９３
• 6.3.2 Preprocessing ９４
• 6.3.3 Hyperparameter Settings ９５
• 6.3.4 Ensemble of Models ９６
• 6.3.5 Evaluation Metrics ９６
• 6.4 Model Interpretability Methodology ９７
• 6.4.1 Attention Weight Analysis ９７
• 6.4.2 Gradient-Based Saliency Analysis of Glottal Features ９８
• 6.4.3 Grad-CAM++-Based Spectrogram Analysis ９９
• 6.5 Results １０１
• 6.5.1 Experimental Results １０１
• 6.5.2 Model Interpretability Results １０４
• 6.6 Discussion １１０
• 6.6.1 Model Performance １１０
• 6.6.2 Model Interpretation １１４
• 6.7 Summary １２１
• Chapter 7 Conclusion and Future Directions １２３
• 7.1 Summary of Key Findings １２３
• 7.2 Dissertation Contributions １２５
• 7.3 Limitations and Future Research Directions １２７
• Bibliography １３２
• Appendix A Speech Stimuli １５８
• Appendix B Group Comparison: Control vs. Patients １５９
• Appendix C Group Comparison: CKD stages １６５
• Appendix D Correlation Analysis １７４
• Abstract in Korean １７７
• Acknowledgments １７９