本研究旨在评估源自单个和多个肿瘤部位的治疗前基于 CT 的放射组学特征 (RF) 以及最先进的机器学习生存算法在预测无进展生存期 (PFS) 方面的潜力适用于接受免疫检查点抑制剂（CPI）等一线治疗的转移性肺腺癌（MLUAD）患者。为此，纳入了 2016 年 11 月至 2022 年 11 月期间在我们的癌症中心接受一线 CPI 治疗的所有新诊断 MLUAD、治疗前对比增强 CT 扫描且体能状态≤ 2 的成人。从 CT 扫描中体积≥1 cm3 的所有可测量病灶中提取 RF。为了捕获肿瘤内和肿瘤间的异质性，收集了每位患者最大肿瘤的 RF 以及每位患者所有病变的最低、最高和平均 RF 值。计算患者内肿瘤间异质性指标以测量每个患者病变之间的相似性。在过滤单变量 Cox p < 0.100 的预测变量并分析其相关性后，使用五种生存机器学习算法（逐步 Cox 回归 [SCR]、LASSO Cox 回归、随机生存森林、梯度提升机 [GBM] 和深度学习 [Deepsurv]）接受 100 次重复 5 倍交叉验证 (rCV) 训练，以根据三个输入预测 PFS：(i) 临床病理学变量，(ii) 所有基于放射组学和临床病理学（完整输入），以及 (iii) 不相关的放射组学 -基于变量和临床病理学变量（不相关的输入）。使用一致性指数（c 指数）评估模型的性能。总体而言，纳入了 140 名患者（中位年龄：62.5 岁，36.4% 为女性）。在 rCV 中，Deepsurv 达到最高 c 指数（c 指数 = 0.631，95%CI = 0.625-0.647），其次是 GBM（c 指数 = 0.603，95%CI = 0.557-0.646），显着优于标准SCR 无论其输入如何（c 指数范围：0.560-0.570，所有 p < 0.0001）。因此，当使用先进的机器学习生存算法进行分析时，单部位和多部位治疗前放射组学数据为预测接受一线 CPI 治疗的 MLUAD 患者的 PFS 提供了有价值的预后信息。

This study aimed to evaluate the potential of pre-treatment CT-based radiomics features (RFs) derived from single and multiple tumor sites, and state-of-the-art machine-learning survival algorithms, in predicting progression-free survival (PFS) for patients with metastatic lung adenocarcinoma (MLUAD) receiving first-line treatment including immune checkpoint inhibitors (CPIs). To do so, all adults with newly diagnosed MLUAD, pre-treatment contrast-enhanced CT scan, and performance status ≤ 2 who were treated at our cancer center with first-line CPI between November 2016 and November 2022 were included. RFs were extracted from all measurable lesions with a volume ≥ 1 cm3 on the CT scan. To capture intra- and inter-tumor heterogeneity, RFs from the largest tumor of each patient, as well as lowest, highest, and average RF values over all lesions per patient were collected. Intra-patient inter-tumor heterogeneity metrics were calculated to measure the similarity between each patient lesions. After filtering predictors with univariable Cox p < 0.100 and analyzing their correlations, five survival machine-learning algorithms (stepwise Cox regression [SCR], LASSO Cox regression, random survival forests, gradient boosted machine [GBM], and deep learning [Deepsurv]) were trained in 100-times repeated 5-fold cross-validation (rCV) to predict PFS on three inputs: (i) clinicopathological variables, (ii) all radiomics-based and clinicopathological (full input), and (iii) uncorrelated radiomics-based and clinicopathological variables (uncorrelated input). The Models' performances were evaluated using the concordance index (c-index). Overall, 140 patients were included (median age: 62.5 years, 36.4% women). In rCV, the highest c-index was reached with Deepsurv (c-index = 0.631, 95%CI = 0.625-0.647), followed by GBM (c-index = 0.603, 95%CI = 0.557-0.646), significantly outperforming standard SCR whatever its input (c-index range: 0.560-0.570, all p < 0.0001). Thus, single- and multi-site pre-treatment radiomics data provide valuable prognostic information for predicting PFS in MLUAD patients undergoing first-line CPI treatment when analyzed with advanced machine-learning survival algorithms.