Background: This study addresses the significant challenge of low survival rates in patients with cause-specific lung cancer accompanied by bone or brain metastases. Recognizing the critical need for an effective predictive model, the research aims to establish survival prediction models using both parametric and non-parametric approaches. Methods: Clinical data from lung cancer patients with at least one bone or brain metastasis between 2000 and 2020 from the SEER database were utilized. Four models were constructed: Cox proportional hazard, Weibull accelerated failure time (AFT), log-normal AFT, and Zografos–Balakrishnan log-normal (ZBLN). Independent prognostic factors for cause-specific survival were identified, and model fit was evaluated using Akaike’s and Bayesian information criteria. Internal validation assessed predictive accuracy and discriminability through the Harriel Concordance Index (C-index) and calibration plots. Results: A total of 20,412 patients were included, with 14,290 (70%) as the training cohort and 6122 (30%) validation. Independent prognostic factors selected for the study were age, race, sex, primary tumor site, disease grade, total malignant tumor in situ, metastases, treatment modality, and histology. Among the accelerated failure time (AFT) models considered, the ZBLN distribution exhibited the most robust model fit for the 3- and 5-year survival, as evidenced by the lowest values of Akaike’s information criterion of 6322 and 79,396, and the Bayesian information criterion of 63,495 and 79,396, respectively. This outperformed other AFT and Cox models (AIC = [156,891, 211,125]; BIC = [158,848, 211,287]). Regarding predictive accuracy, the ZBLN AFT model achieved the highest concordance C-index (0.682, 0.667), a better performance than the Cox model (0.669, 0.643). The calibration curves of the ZBLN AFT model demonstrated a high degree of concordance between actual and predicted values. All variables considered in this study demonstrated significance at the 0.05 level for the ZBLN AFT model. However, differences emerged in the significant variations in survival times between subgroups. The study revealed that patients with only bone metastases have a higher chance of survival compared to only brain and those with bone and brain metastases. Conclusions: The study highlights the underutilized but accurate nature of the accelerated failure time model in predicting lung cancer survival and identifying prognostic factors. These findings have implications for individualized clinical decisions, indicating the potential for screening and professional care of lung cancer patients with at least one bone or brain metastasis in the future.
背景：本研究针对特定病因肺癌伴骨或脑转移患者生存率低这一重大挑战。认识到建立有效预测模型的迫切需求，本研究旨在运用参数与非参数方法构建生存预测模型。方法：利用SEER数据库中2000年至2020年间至少发生一处骨或脑转移的肺癌患者临床数据，构建了四种模型：Cox比例风险模型、威布尔加速失效时间模型、对数正态加速失效时间模型及Zografos–Balakrishnan对数正态模型。研究确定了影响病因特异性生存的独立预后因素，并采用赤池信息准则和贝叶斯信息准则评估模型拟合度。通过Harrell一致性指数和校准曲线进行内部验证，评估预测准确性与区分能力。结果：共纳入20,412例患者，其中训练队列14,290例（70%），验证队列6,122例（30%）。研究筛选出的独立预后因素包括年龄、种族、性别、原发肿瘤部位、疾病分级、恶性肿瘤总数、转移情况、治疗方式和组织学类型。在考察的加速失效时间模型中，ZBLN分布在3年及5年生存期预测中表现出最优模型拟合度：其赤池信息准则值（6,322和79,396）与贝叶斯信息准则值（63,495和79,396）均为最低，显著优于其他加速失效时间模型及Cox模型（赤池信息准则=[156,891, 211,125]；贝叶斯信息准则=[158,848, 211,287]）。在预测准确性方面，ZBLN加速失效时间模型获得最高一致性指数（0.682, 0.667），优于Cox模型（0.669, 0.643）。其校准曲线显示实际观测值与预测值高度吻合。本研究所有变量在ZBLN加速失效时间模型中均达到0.05水平显著性，但亚组间生存时间存在显著差异。研究发现，相较于单纯脑转移及骨脑双转移患者，单纯骨转移患者生存机会更高。结论：本研究揭示了加速失效时间模型在预测肺癌生存期和识别预后因素方面虽未充分应用但具有精确性的特点。这些发现对个体化临床决策具有指导意义，为未来对至少伴有一处骨或脑转移的肺癌患者开展筛查和专业护理提供了潜在可能。
肿瘤(癌症)患者之家