目的：在我国，结外鼻型NK/T细胞淋巴瘤 (extranodal nasal-type NK/T-cell lymphoma, ENKTCL) 约占全部外周T细胞淋巴瘤患者的40-50%，是最常见的一种亚型。ENKTCL是一类异质性疾病，具有不同的临床特征和预后。基于现代放疗技术（如IMRT）和非蒽环类化疗药物的现代治疗模式，ENKTCL患者的预后有所改善，但疾病复发率仍较高，出现复发进展的患者预后很差。目前已有的ENKTCL预后模型多为评分系统和列线图形式，仅对特定时间点的结局进行预测（如3年或5年总生存率），缺少对复发进展后死亡率的评估，在预测结果的连续性展示上也有所欠缺。本研究将构建ENKTCL多态预测模型，增加对复发进展状态的评估，同时通过交互式界面实现模型的个体化预测和治疗方案选择。方法：对中国淋巴瘤协作组 (China Lymphoma Collaborative Group, CLCG) 的多中心队列进行回顾性分析，构建多态预后模型。模型由三个状态构成：初治、复发进展和死亡状态，在每个状态过渡阶段通过BeSS (Best Subset Selection in Linear, Logistic and CoxPH Models) 方法进行变量筛选。使用多因素Cox回归分析的方法进行建模。通过校准曲线、曲线下面积（Area Under Curve, AUC）、时依AUC（time-dependent AUC）、C指数和决策曲线分析法进行模型评估。最后使用交互式界面进行模型展示。结果：ENKTCL多态预测模型所选预测变量为年龄、LDH（lactate dehydrogenase）、ECOG（Eastern Cooperative Oncology Group）评分、Ann Arbor分期、B症状、PTI、治疗方式、复发进展时间。在各状态下进行Cox回归分析，对于初治状态到复发进展，LDH升高、Ann Arbor分期（II、III、IV）、有原发肿瘤浸润（PTI）均与复发进展风险相关。在初治状态到死亡阶段，年龄、LDH升高、Ann Arbor分期（II和IV）、B症状和ECOG评分≥2分均与死亡风险相关。特别地，复发进展时间对于复发进展后死亡产生显著影响（HR: 1.65; 95%CI: 1.36-2.00），而且在加入治疗方式交互作用后，不同治疗方式的患者复发进展时间依然与复发进展后的死亡风险增高相关。模型评估方面，本模型的校准曲线显示，5年总生存率预测值和实际值之间具有极好的一致性，5年无进展生存率预测同样较好。与现有预后模型，如IPI（International Prognostic Index）、KPI（Korean Prognostic Index）、PINK（prognostic index of natural killer lymphoma）、Nomogram预后模型和NRI（nomogram-revised risk index）相比，本模型的曲线下面积（AUC）、时依AUC和C指数均更高，决策曲线分析结果更优。结论：ENKTCL多态预后模型首次考虑了复发进展后的死亡率，同时该模型具有良好的预测准确性和临床效用，交互式形式可以进行个体化预测，有利于实现个体化医疗。

Purpose: Extranodal natural killer/T-cell lymphoma, nasal-type (ENKTCL) accouts for 40-50% of all peripheral T-cell lymphoma in China. ENKTCL is a heterogeneous disease with the different clinical course and prognosis. The prognosis of ENKTCL has improved by using the modern radiotherapy technology (such as IMRT) and non-anthracycline chemotherapy. However, patients still have a high rate of recurrence and the prognosis of these patients with recurrent disease is extremely poor. At present, most of the ENKTCL prognosic models are in the form of scoring system and nomogram. These prognosic models could only predict the outcome of specific time points, such as 3-year or 5-year overall survival, lacking the evaluation of mortality after recurrence and progression. In this study, we will build a multistate prognostic model of ENKTCL to evaluate the recurrence probability, make individual predictions and recommend the best treatment options through the interactive graphical user interface. Method: The clinical data from the China Lymphoma Collaborative Group (CLCG) was analyzed retrospectively to construct a multistate prognostic model. The model consists of three states: initial treatment, recurrence progress and death. On each transition state, the Best Subset Selection in Linear, Logistic and CoxPH Models (BeSS) is used to find variables. The multivariate Cox regression analysis is used to build this model. The model is evaluated by calibration curve, area under curve (AUC), time-dependent AUC, the Harrell’s C-index and decision curve analysis. Finally, the interactive graphical user interface is used to provide risk estimates from underlying multistate prediction model after user inputs predictor values. Result: The predicted variables of multistate prognostic model included age, lactate dehydrogenase (LDH), Eastern Cooperative Oncology Group (ECOG) score, Ann Arbor stage, B symptom, primary tumor invasion (PTI), different therapies and recurrent time. Cox regression analysis showed that elevated LDH, Ann Arbor stage (II and III/IV) and PTI were associated with the risk of recurrence and progression on the transition from treatment to recurence. On the transition from treatment to death，age, elevated LDH, ECOG score ≥ 2, Ann Arbor stage (II and III/IV) and B symptom were all related to the risk of death. In particular, the recurrent time had a significant impact on the mortality after disease recurrence (HR: 1.65; 95% CI: 1.36-2.00). Moreover, the recurrent time of patients with different therapies was still associated with the increased risk of death after disease recurrence under the covariate-treatment interactions. As for model evaluation, the calibration curve of the multistate prognostic model showed excellent consistency between predicted and actual values of 5-year overall survival (OS) and 5-year progression free survival (PFS). Compared with the IPI (International Prognostic Index), KPI (Korean Prognostic Index), PINK (prognostic index of natural killer lymphoma), nomogram prognostic model and NRI (nomogram-revised risk index), the AUC, time-dependent AUC and Harrell’s C-index of the multistate prognostic model was higher and decision curve analysis also showed a superior outcome. Conclusion: The multistate prognostic model of ENKTCL is the first study to incorporate mortality after disease recurrence. Meanwhile, this multistate model has better prediction accuracy and clinical utility. Interactive graphical user interface can be used for individual prediction, which is conducive to the individualized prediction and treatment options.