비부비동 악성종양 치료 예후 예측을 위한 대규모 언어모델 기반 머신러닝 연구 = Large Language Model-Assisted Machine Learning for Treatment Outcome Prediction in Sinonasal Malignancies

다국어 초록 (Multilingual Abstract)

Objective This study aimed to develop and evaluate machine learning (ML) models for predicting survival in patients with rare and heterogeneous sinonasal malignancies, identify significant prognostic factors, and compare their performance with traditional statistical methods. The secondary objectives were to extensively integrate large language models (LLMs) into the ML pipeline to assess their feasibility in reliably building ML models and to explore their practical applicability in medical research. Methods We retrospectively analyzed a cohort of 621 patients with sinonasal malignancies treated at a single tertiary center between 1990 and 2024. LLMs were incorporated into the analytic workflow for data preprocessing, feature engineering, and the development of multiple ML models, including logistic regression, random forest, XGBoost, LightGBM, support vector classifier (linear SVC), and multilayer perceptron for 5-year survival classification, as well as Cox proportional hazards regression, random survival forest, XGBoost-Survival, LightGBM-Survival, and DeepSurv for survival analysis. All the steps of model training, hyperparameter tuning, and variable visualization were initially attempted within the LLM environment; when execution was not feasible, the code generated by the LLM was externally verified and run by investigators. Subgroup analyses were performed for squamous cell carcinoma and epithelial and non-epithelial tumors, and additional models were constructed for recurrence prediction. Results For 5-year survival prediction in the overall cohort, random forest achieved the highest area under the receiver operating characteristic curve (AUROC) on the V1 set (0.861), while LightGBM yielded the best F1-score (0.718). The V2 feature set notably improved linear models, with the AUROC of logistic regression increasing to 0.854 (Δ +0.025). In the subgroup analyses, the performance was particularly strong in SCC, where the V2 set increased the AUROC of logistic regression to 0.906. For the time-to-death analysis, the tuned XGBoost-Survival model demonstrated superior performance (C-index 0.777), significantly outperforming the traditional Cox proportional hazards model (C-index 0.752). Across diverse models and subgroups, residual disease after initial treatment, distant metastasis, and age at diagnosis consistently emerged as the most powerful prognostic factors. The LLM-integrated workflow proved feasible; however, certain models including LightGBM and all survival analysis algorithms required external execution due to library constraints. Conclusions ML models provided more accurate survival prediction than conventional methods and consistently identified key prognostic factors including residual disease, distant metastasis, and age. By integrating LLMs into the analytic workflow, this study presents a practical framework that links advanced computational methods with clinical applicability, enhancing prognostic accuracy and supporting the adoption of AI-based prognostic modeling for rare sinonasal cancers in otolaryngology.

목차 (Table of Contents)

• I. 서론 1
• II. 연구대상 및 방법 3
• 1. 계산 환경 및 대규모 언어모델 통합 과정 3
• 2. 연구 설계 및 데이터 수집 3
• 3. 데이터 전처리 5
• 4. 특징 엔지니어링 및 선택 6
• 5. 모델 개발 7
• 6. 모델 평가 7
• 7. 하위 그룹 분석 8
• 8. 대규모 언어모델 활용 범위 9
• A. 데이터 전처리 9
• B. 모델 개발 10
• C. 모델 해석 및 시각화 11
• D. 프롬프트 설계 12
• 9. 검증 과정 13
• III. 연구 결과 15
• 1. 환자군의 인구학적 및 임상적 특성 15
• 2. 모델 성능 15
• A. 이진 분류 15
• B. 생존 분석 16
• C. 모델 해석 17
• 3. 대규모 언어 모델의 성능 및 한계 18
• A. 모델 개발 18
• B. 오류 처리 및 코드 개선 19
• C. 모델 해석 및 시각화 19
• IV. 고찰 20
• 1. 주요 결과 요약 20
• 2. 기계 학습 모델의 임상적 의미- 20
• 3. 기계학습 모델로 확인된 비부비동 악성종양의 주요 예후 인자 23
• A. 초기 치료 후 잔존 종양 23
• B. 원격 전이 24
• C. 고령 25
• D. 수술 25
• E. 추적 관찰 중 재발 26
• F. 진단 시기 26
• G. NLR 27
• H. 기타 변수 28
• 4. 대규모 언어모델 통합의 역할 및 실제 적용 가능성 29
• A. 기계학습 전 과정에 걸친 대규모 언어모델의 기여 30
• B. 연구자의 기술적 장벽을 낮출 수 있는 잠재력 31
• C. 대규모 언어모델의 한계와 연구자의 역할 32
• 5. 연구의 한계 34
• V. 결론 35
• 참고문헌 75
• 부록 1: V1 데이터셋에 포함된 전체 변수 목록 79
• 부록 2: 변수 선택 방법 및 예측 과제별 V2 변수셋 80
• 부록 3: 최적 하이퍼파라미터 및 모델 성능 81
• 부록 4: 실제 실행 파이썬 코드 83
• 부록 5: 사용 소프트웨어 및 환경 정보 89
• 부록 6: 프롬프트 예시 90
• 부록 7: 모델별 상위 중요 변수 91
• 국문요약 92