Abstract: Background In recent years，immune checkpoint inhibitors（ICIs） have shown remarkable efficacy in lung cancer treatment，but their immune-related adverse events（irAEs） have also received widespread attention. Thyroid injury is the most common endocrine irAEs，so it is of great clinical value to construct a prediction model. Objective To establish a prediction model of thyroid irAE in lung cancer patients receiving ICI therapy. Methods A total of 243 lung cancer patients treated with ICI in Hebei Provincial People's Hospital from January 2020 to March 2024 were retrospectively included as study subjects，and randomly divided into training set（169 cases） and validation set（74 cases） according to a ratio of 7 ︰ 3. According to thyroid function，the training set was divided into thyroid irAE group（71 cases） and thyroid non-irAE group（98 cases）. General data and laboratory test indicators of the patients were collected，and subgroup analysis was performed according to Common Terminology Criteria for Adverse Events（CTCAE）. Univariate analysis was performed to screen variables，and multivariate logistic regression was used to analyze the independent influencing factors of thyroid irAE. Variance inflation factor was used to evaluate multicollinearity among the predictors. A thyroid irAE nomogram model was constructed based on the results of multivariate logistic regression analysis. The receiver operating characteristic（ROC） curve of thyroid irAE was drawn，and the area under ROC curve（AUC） was calculated. The model was validated internally by Bootstrap self-sampling method，and the model efficacy was evaluated by Hosmer Lemeshow test，decision curve analysis（DCA），and clinical impact curve（CIC）. Kaplan-Meier analysis was used to compare the cumulative incidence of thyroid irAE in each risk stratification. Multiple logistic regression was used to analyze the influencing factors of thyroid irAE in patients with different CTCAE grades. Results Among 169 patients with lung cancer，138（81.66%） were males and 31（18.34%） were females. The median age was 66 years（60，71），71 cases（42.01%） in thyroid irAE group and 98 cases（57.99%） in thyroid non-irAE group. There were significant differences in ICI treatment cycle，Ki-67，tumor size and FT3 between the two groups（P<0.05）. Multivariate Logistic regression analysis showed that TSH（OR=1.636，95%CI=1.070-2.503，P=0.023），FT3（OR=6.868，95%CI=2.812-16.776，P<0.001），tumor size（OR=0.965，95%CI=0.942-0.989，P=0.004），Ki-67（OR=1.028，95%CI=1.008-1.048，P=0.005） and CYFRA21-1（OR=1.050，95%CI=1.016-1.085，P=0.003） were the independent influencing factors for thyroid irAE in lung cancer patients（P<0.05）. Using TSH，FT3，Ki- 67，CYFRA21-1 and tumor size as predictors，a nomogram model was constructed. The AUC，specificity and sensitivity of thyroid irAE were 0.796，0.827 and 0.634 in the diagnosis of lung cancer patients by nomogram model. The validation set AUC was 0.730，the specificity was 0.810，and the sensitivity was 0.562. The C-index calculated by Bootstrap internal verification is 0.796，and the accuracy rate is 75%，indicating that the model has a good degree of differentiation. Hosmer-Lemeshow test results showed that the model had a good fit（P>0.05）. The calibration curve showed that the predicted probability of thyroid irAE was in good agreement with the observed value. The DCA curve shows that the model can provide a good clinical net benefit in the probability range of 0 to 95%. The CIC curve showed that with the increase of the model threshold，the predicted occurrence of thyroid irAE was consistent with the actual diagnosis results，indicating that the model had good clinical practicability. Based on the optimal probability threshold of thyroid irAE prediction model（Pr value ≥ 0.474），243 patients with lung cancer were divided into high-risk group 94（38.68%）and low-risk group 149（61.32%）. Kakaplan Meier analysis showed that the cumulative incidence of thyroid irAE in high-risk group was significantly higher than that in low-risk group at the 6th treatment cycle（χ2 =28.15，P<0.001），and the incidence risk in high-risk group was 2.63 times higher than that in low-risk group（HR=2.63，95%CI=1.770-3.905，P<0.001）. The results of multivariate logistic analysis of influencing factors of thyroid irAE at different CTCAE grades showed that FT3（OR=5.513，95%CI=1.846-16.465，P=0.002），tumor size（OR=0.963，95%CI=0.928-0.999，P=0.044） were the independent influencing factors of thyroid irAE in CTCAE ≥ 2 subgroups. TSH，FT3，tumor size，Ki-67 and CYFRA21-1 were the independent influencing factors of thyroid irAE in CTCAE grade 1 subgroup（P<0.05），suggesting that the model has more predictive advantages for thyroid irAE of CTCAE grade 1. Conclusion In this study，thyroid irAE prediction model was established based on TSH，FT3，tumor size，Ki-67 and CYFRA21-1. Active monitoring of thyroid function in patients with prediction probability ≥ 47.4% is helpful to reduce the risk of immunotoxicity and improve the quality of life.

Key words: Lung cancer, Immune checkpoint inhibitors, Thyroid immune-related adverse event, Prediction model

摘要： 背景 近年来，免疫检查点抑制剂（ICIs）在肺癌治疗中展示出显著疗效，但其引发的免疫相关不良事件（irAEs）也受到广泛关注。甲状腺损伤作为最常见的内分泌 irAEs，构建预测模型具有重要临床价值。目的 构建肺癌患者ICI治疗后甲状腺irAE的预测模型。方法 回顾性纳入2020年1月—2024年3月在河北省人民医院接受ICI 治疗的 243 例肺癌患者为研究对象，按照 7 ∶ 3 的比例进行随机抽样分为训练集（n=169）和验证集（n=74）。根 据甲状腺功能将训练集分为甲状腺 irAE 组（n=71）和无甲状腺 irAE 组（n=98），收集患者一般资料及实验室检查指标，并根据常见不良事件评价标准（CTCAE）进行亚组分析。采用单因素分析筛选变量，并通过多因素 Logistic 回归分析甲状腺 irAE 的独立影响因素；采用方差膨胀因子评估预测因子之间多重共线性。基于多因素 Logistic 回归分析结果构建甲状腺 irAE 列线图模型，绘制模型预测甲状腺 irAE 的受试者工作特征（ROC）曲线，并计算 ROC 曲线下面积（AUC）；采用 Bootstrap 自助抽样法对模型进行内部验证，并通过 Hosmer-Lemeshow 检验、决策曲线分析（DCA）、临床影响曲线（CIC）评估模型效能。通过 Kaplan-Meier 分析比较各风险分层甲状腺 irAE 的累积发生率。采用多因素 Logistic 回归分析不同 CTCAE 分级患者发生甲状腺 irAE 影响因素。结果 训练集 169 例肺癌患者中，男性 138 例（81.66%），女性 31 例（18.34%）；中位年龄 66 岁（60，71），甲状腺 irAE 组 71 例（42.01%），无甲状腺 irAE 组 98例（57.99%）。两组 ICI 治疗周期、Ki-67、肿瘤大小、FT3 比较，差异有统计学意义（P<0.05）。多因素 Logistic 回归分析 结 果 显 示，TSH（OR=1.636，95%CI=1.070~2.503，P=0.023）、FT3（OR=6.868，95%CI=2.812~16.776，P<0.001）、 肿瘤 大 小（OR=0.965，95%CI=0.942~0.989，P=0.004）、Ki-67（OR=1.028，95%CI=1.008~1.048，P=0.005）、CYFRA21-1（OR=1.050，95%CI=1.016~1.085，P=0.003）是肺癌患者发生甲状腺 irAE 的独立影响因素（P<0.05）。以 TSH、FT3、Ki-67、CYFRA21-1 及肿瘤大小为预测因子构建列线图模型，列线图模型诊断肺癌患者发生甲状腺 irAE 的 AUC 为 0.796，特异度为 0.827，灵敏度为 0.634；验证集 AUC 为 0.730，特异度为 0.810，灵敏度为 0.562。Bootstrap 内部验证计算得 C-index为 0.796，准确率 75%，表明模型具有良好的区分度；Hosmer-Lemeshow 检验结果表明模型拟合度良好（P>0.05）。校准曲线显示甲状腺 irAE 发生的预测概率与实际观测值一致良好；DCA 曲线显示在 0~95% 的概率范围内，该模型可提供良好临床净获益；CIC 曲线表明随着模型阈值的增加，预测发生甲状腺 irAE 与实际诊断结果一致，表明模型临床实用性好。基于甲状腺 irAE 预测模型最佳概率阈值（Pr 值≥ 0.474），将 243 例肺癌患者分为高危组 94 例（38.68%）和低危组 149例（61.32%）。Kaplan-Meier 分析结果显示，第 6 治疗周期时高危组甲状腺 irAE 累积发生率显著高于低危组（χ2 =28.15，P<0.001），高危组发病风险是低危组 2.63 倍（HR=2.63，95%CI=1.770~3.905，P<0.001）。不同 CTCAE 分级甲状腺 irAE影响因素的多因素 Logistic 分析结果显示，FT3（OR=5.513，95%CI=1.846~16.465，P=0.002）、肿瘤大小（OR=0.963，95%CI=0.928~0.999，P=0.044）是 CTCAE ≥ 2 级亚组甲状腺 irAE 的独立影响因素；而 TSH、FT3、肿瘤大小、Ki-67 及CYFRA21-1 均是 CTCAE 1 级亚组甲状腺 irAE 的独立影响因素（P<0.05），提示模型对 CTCAE 1 级甲状腺 irAE 更具预测优势。结论 本研究基于 TSH、FT3、肿瘤大小、Ki-67 及 CYFRA21-1 建立了甲状腺 irAE 预测模型，预测概率≥ 47.4%者积极监测甲状腺功能有助于降低免疫毒性风险、提高生存质量。

关键词: 肺癌, 免疫检查点抑制剂, 甲状腺免疫相关不良反应, 预测模型