鼻咽癌病患台灣本土預後模型開發與跨族群驗證

Abstract

研究背景與目的 鼻咽癌在流行病學上有顯著的地理分佈差異,尤其在東亞與東南亞地區發生率較高。雖然傳統的美國聯合癌症聯合委員會 (AJCC) 分期系統提供了統一的臨床預後評估標準,但其主要基於解剖學範圍,未能納入臨床病理特徵及生活習慣等重要預後因子,因此本研究旨在開發一個整合臨床病理參數與生活行為因子的多變項預測模型,並透過內外部驗證提升鼻咽癌患者在整體存活率、癌症特異性存活率及無病存活率上的精準度。 研究方法 本研究利用台灣癌症登記資料庫 (TCR) 進行模型開發與內部驗證。主要建構兩個不同的多變項 Cox 比例風險模型:模型 1 為單變項顯著且可透過美國監測、流行病學及最終結果計畫 (SEER) 驗證的參數,包含年齡、性別、T 與 N分期、腫瘤分級及組織學類型;模型 2 則進一步納入 BMI、抽菸、飲酒及嚼食檳榔等行為因子。主要透過 10 倍交叉進行內部驗證,並利用 SEER 資料庫針對模型 1 進行跨種族(亞裔、白人、非裔)的外部驗證,以 Harrell’s C-index 與校正分析評估模型表現。 研究結果 本研究模型在各個存活預測指標上均有穩定精準的辨別力。在內部驗證中,模型 1 與模型 2 的整體存活率 C-index 分別為 0.7152 與 0.7155;癌症特異性存活率也都達到 0.7174,校正分析也顯示預測事件數與實際觀察值具高度一致性;外部驗證進一步證實了該框架在亞洲族群中具備良好的預測外推性。相較於單純使用 AJCC 分期系統在各存活終點的預測局限,其 C-index 僅介於0.613 至 0.647 之間,本研究模型提供了實質的預測增益。 結論 本研究整合臨床病理特徵,建立了一個針對台灣族群的鼻咽癌預後預測模型,相較傳統 AJCC 分期能提供更精準的風險分層,且具外部驗證效力,能為臨床決策提供實質建議與價值。

Background and Objectives The epidemiology of nasopharyngeal carcinoma (NPC) is characterized by a remarkable geographic heterogeneity, demonstrating a disproportionately high prevalence within East and Southeast Asian populations. Although the conventional AJCC staging system is a cornerstone for clinical prognosis, its primary focus on anatomical extent overlooks critical prognostic factors such as clinicopathological characteristics and lifestyle behaviors. This study aimed to develop a multivariable prognostic framework integrating clinicopathological and lifestyle variables. By employing both internal and external validation, the research sought to enhance predictive accuracy for overall survival (OS), cancer-specific survival (CSS), and disease-free survival (DFS) in patients with NPC. Methods The Taiwan Cancer Registry (TCR) was utilized for model development and internal validation. Two distinct multivariable Cox proportional hazards models were constructed: Model 1 included univariate significant factors verifiable within the Surveillance, Epidemiology, and End Results (SEER) Program, specifically age, sex, T and N categories, tumor grade, and histological type; Model 2 further incorporated BMI and lifestyle factors, including smoking, alcohol consumption, and betel nut chewing. Internal validation was performed using 10-fold cross-validation, while external validation of Model 1 was conducted across racial subgroups (Asian, White, and Black) using the SEER database. Model performance across both validation phases was evaluated using Harrell’s C-index and calibration analysis to assess discriminative ability and reliability. Results The integrated models demonstrated robust discriminative ability across all survival endpoints. In the internal validation, the C-indices for OS reached 0.7152 in Model 1 and 0.7155 in Model 2, while both models achieved a C-index of 0.7174 for CSS. Calibration analysis confirmed a high degree of agreement between predicted events and actual observed outcomes. External validation supported the generalizability of the models to Asian populations. Compared to the prognostic limitations of the conventional AJCC staging system alone across survival endpoints, which provided C-index ranging modestly between 0.613 and 0.647, this integrated framework delivered a substantial predictive increment. Conclusions This study established a prognostic prediction model for NPC tailored to the Taiwanese population by integrating diverse clinical variables. Compared to the traditional AJCC staging system, this framework provides more refined risk stratification and possesses strong external validation efficacy, offering substantial recommendations and practical value for clinical decision-making.