Development of an Efficient Privacy-Preserving U-Shape Inference Framework for Bladder Cancer Surveillance Using Urine Proteomic Biomarkers and Homomorphic Encryption = 소변 단백체 바이오마커와 동형암호 기반의 효율적인 프라이버시 보존 U-자형 방광암 감시 추론 프레임워크 개발

다국어 초록 (Multilingual Abstract)

Purpose: This study aimed to develop an efficient privacy-preserving U-shape inference framework for bladder cancer surveillance using urine proteomic biomarkers. The framework divides encrypted computation between client and server, adopting split inference to balance diagnostic accuracy, computational efficiency, and data confidentiality.
Methods: Urine proteomic profiling identified 676 exosomal and 935 soluble proteins, of which 132 and 248 were differentially expressed between cancer and control samples. Eight proteins—A2M, AFM, APOA1, CFL1, ITIH2, FGB, CDC5L, and CD5L—were validated by enzyme-linked immunosorbent assay in 120 urine samples (95 urothelial carcinoma, 25 benign). Two logistic-regression models were reconstructed: a three-protein model (ITIH2, AFM, CFL1) and an eight-protein model. Both were validated using stratified three-fold cross-validation with bootstrap 95 percent confidence intervals for AUROC and AUPRC. Encrypted inference was implemented with the CKKS scheme (polynomial modulus 16 384; coefficient-modulus chain 60, 40, 40, 60 bits; scale 2⁴⁰). In the U-shape design, the client encrypts input data, the server performs ciphertext–plaintext dot-product operations with encrypted inputs and model weights and returns the encrypted linear score for client-side decryption and sigmoid computation. A fully encrypted sigmoid model using Chebyshev approximation was tested for comparison. Latency, communication, and numerical deviation, defined as the absolute difference between encrypted and plaintext probabilities, were measured.
Results: The three- and eight-protein models achieved AUROC values of 0.810 (95 % CI, 0.701–0.904) and 0.806 (0.709–0.893) with AUPRC above 0.93, demonstrating stable discrimination across both panels. The three-protein model was selected for encrypted inference implementation due to its reduced computational complexity. Split inference reproduced plaintext results with a mean absolute probability difference of 5.6 × 10⁻⁵ and an end-to-end latency of 63 milliseconds per sample (15 milliseconds for encryption, 44 milliseconds for the ciphertext–plaintext dot-product, 4 milliseconds for decryption) and a communication load of 7 megabytes. The fully encrypted sigmoid evaluation added 106 milliseconds of polynomial computation, increasing total latency to 169 milliseconds per sample and raising the mean absolute probability error to 0.165, highlighting that substantial optimization of the polynomial approximation is required to achieve clinical-grade accuracy. The split U-shape framework was therefore adopted, maintaining diagnostic fidelity with minimal computational overhead and dual protection of patient data and model intellectual property.
Conclusions: The U-shape privacy-preserving inference framework enables secure encrypted inference without degradation of diagnostic accuracy. By executing linear computations on the server and nonlinear inference on the client, it preserves privacy and model confidentiality while maintaining latency suitable for real-time bladder cancer surveillance.

국문 초록 (Abstract)

서론: 본 연구는 소변 단백체 바이오마커를 기반으로 방광암을 비침습적으로 감시할 수 있는 진단 모델을 개발하고, 이를 개인정보 보호하며 활용하기 위한 프라이버시 보존 U-자형(split) 분할 추론 프레임워크를 구현하고자 하였다.

방법: 질량분석 기반 단백체 분석을 통해 후보 단백질을 선별하고, A2M, AFM, APOA1, CFL1, ITIH2, FGB, CDC5L, CD5L의 8개 단백질을 120개 소변 샘플(방광암 95건, 양성 25건)을 대상으로 ELISA로 검증하였다. 검증된 데이터를 이용해 3단백질 모델(ITIH2, AFM, CFL1)과 8단백질 모델(8종 전체)을 구성하여 로지스틱 회귀 기반 진단 모델을 구축하였다. 모델은 층화 3겹 교차검증과 부트스트랩을 통해 성능을 평가하였으며, AUROC, AUPRC, 민감도, 특이도, NPV, PPV를 분석하였다. 암호화 추론은 CKKS 동형암호 스킴을 기반으로 구현하였다. 클라이언트는 데이터를 암호화하여 서버로 전송하고, 서버는 암호문–평문 내적 연산을 수행한 후 암호화된 결과를 반환하였으며, 클라이언트는 복호화 후 시그모이드 변환을 수행하는 U-자형 구조를 적용하였다. 또한 시그모이드 변환을 근사다항식으로 변환하여 완전동형암호(FHE)로 처리하는 비교 모델을 구현해 성능, 지연 시간, 통신량을 비교하였다.

결과: 3단백질 및 8단백질 모델의 AUROC는 각각 0.810(95% CI, 0.701–0.904)과 0.806(0.709–0.893)으로 측정되었으며, 두 모델 모두 AUPRC 0.93 이상으로 안정적인 진단 성능을 보였다. 계산 복잡도가 낮은 3단백질 모델을 암호화 추론 구현에 사용하였다. U-자형 분할 추론 프레임워크에서 암호화 추론 결과와 평문 연산 결과 간의 평균 절대 확률 오차는 5.6×10⁻⁵로 나타났으며, 총 지연 시간은 63ms(암호화 15ms, 서버 연산 44ms, 복호화 4ms), 통신량은 약 7MB였다. 완전동형암호(FHE) 모델은 다항식 연산에 106ms가 추가되어 총 지연 시간이 169ms로 증가하였으며, 평균 확률 오차는 0.165로 나타나 임상 수준의 정확도 달성을 위해서는 다항식 근사의 추가 최적화가 필요함을 확인하였다. 제안한 U-자형 분할 추론 방식은 최소한의 계산 부담으로 진단 정확도를 유지하면서 환자 데이터와 모델 지적 재산권을 동시에 보호하였다.

결론: 본 연구의 프라이버시 보존 U-자형 분할 추론 프레임워크는 진단 정확도의 손실 없이 안전하고 효율적인 암호화 추론을 가능하게 함을 입증하였다. 서버에서 선형 연산을, 클라이언트에서 비선형 추론을 수행함으로써 환자 데이터와 모델 기밀성을 보호하면서 실시간 방광암 감시에 적합한 수준의 계산 효율성을 달성하였다.

목차 (Table of Contents)

• Table of Contents
• Chapter 1. Introduction 1
• 1.1. Global Disease Burden and Clinical Significance of Bladder Cancer 1
• 1.2. Anatomical Structure of the Bladder Wall 2
• 1.3. Histological Classification and Disease Categories 3
• 1.4. Risk Factors and Etiology 5
• 1.5. Field Cancerization and Clonal Evolution 6
• 1.6. Genomic Landscape and Shared Genetic Alterations 8
• 1.7. Divergent Molecular Pathways: NMIBC vs. MIBC 9
• 1.8. The Tumor Microenvironment and its Role in Biomarker Shedding 10
• 1.9. Development of Urine Biomarkers 11
• 1.10. Multiplex Biomarker Models for Diagnosis and Surveillance 13
• 1.11. Challenges of Implementation and Data Security in Computational Diagnostics 14
• 1.12. Privacy Enhancing Technologies in Healthcare 16
• 1.13. Proposed U-Shape (Split) Inference Architecture 17
• 1.14. Study Aims 19
• Chapter 2. Methods 20
• 2.1.Urine Protein Biomarker Discovery and Validation 20
• 2.1.1. Ethics Approval and Informed Consent 20
• 2.1.2. Patient Selection and Study Flow 20
• 2.1.3. Urine and Exosome Preparation 22
• 2.1.4. Liquid Chromatography–Tandem Mass Spectrometry (LC–MS/MS) 23
• 2.1.5. Data Processing and Bioinformatics Analysis 23
• 2.1.6. Exosome Validation and Candidate Biomarker Selection 24
• 2.1.7. ELISA Validation of Candidate Proteins 25
• 2.1.8. Model Construction and Statistical Analysis 25
• 2.2. Implementation of the U-shape Split Inference Framework 26
• 2.2.1. Overview of the Framework 26
• 2.2.2. Homomorphic Encryption Configuration 29
• 2.2.3. Split Inference Workflow 29
• 2.2.4. Fully Homomorphic Encryption for Comparison 31
• 2.2.5. Evaluation Metrics 31
• 2.2.6. Statistical Analysis 32
• Chapter 3. Results 33
• 3.1. Patient Characteristics 33
• 3.2. Biomarker Discovery by Label-Free Quantitative Proteomics 37
• 3.3. Biomarker Verification by Data-Independent Acquisition 39
• 3.4. Selection of Candidate Biomarkers 42
• 3.5. Quantitative Validation by ELISA 45
• 3.6. Numerical Fidelity and Diagnostic Consistency 53
• 3.7. Computational Latency Analysis 55
• 3.8. Impact of Polynomial Approximation on Precision 55
• 3.9. Communication and Data Transfer Overhead 57
• Chapter 4. Discussion 59
• 4.1. Overview 59
• 4.2. Clinical Rationale and Unmet Need in Bladder Cancer Surveillance 60
• 4.3. Biological Interpretation of the Identified Biomarkers 61
• 4.4. Rationale for selection of the Minimized Variable Model 63
• 4.5. Choice of Urine Proteins over Exosomal Biomarkers 64
• 4.6. Comparison with Existing Urine Biomarkers 66
• 4.7. Integration into a Privacy-Preserving Computational Framework 68
• 4.8. Comparison with Alternative Privacy-Preserving Approaches 70
• 4.9. Clinical Implementation Considerations 72
• 4.10. Strengths and Limitations 74
• 4.11. Future Perspectives 76
• Chapter 5. Conclusion 78
• Bibliography 79
• Abstract in Korean 99
• List of Tables
• Table 1 34
• Table 2 35
• Table 3 36
• Table 4 43
• Table 5 46
• List of Figures
• Figure 1 21
• Figure 2 28
• Figure 3 38
• Figure 4 40
• Figure 5 41
• Figure 6 48
• Figure 7 50
• Figure 8 52
• Figure 9 54
• Figure 10 56
• Figure 11 58