Background: Conventional prognostic factors are typically assessed at diagnosis in metastatic hormone-sensitive prostate cancer (mHSPC). However, variations in vital signs and laboratory parameters occur during systemic treatment and may predict patients’ prognosis and anticipate organ-specific toxicity development. Methods: This single-center retrospective study included 363 patients with de novo mHSPC treated between 2014 and 2023. Clinical and laboratory data were systematically collected from the hospital data warehouse, from treatment initiation through the following seven months. Variations in vital parameters and blood test results were graded using CTCAE V5.0 (dynamic variables). Cox regression analyses were performed to explore the impact of dynamic variables on progression-free survival (PFS) and overall survival (OS). Machine learning (ML) models (Support Vector Classifier, Random Forest, and LGBM Classifier) were developed to predict single organ-specific toxicities and to identify good and poor responders based on 7-month PSA levels, PFS and OS. We compared ML model performance when trained only on baseline factors (static models) with those integrating variables generated by vital sign and blood test monitoring within 3 and 7 months from treatment start (dynamic models). Results: Dynamic model failed to improve the prediction of single organ-specific toxicities. Univariable Cox analysis revealed that the development of hematological, liver, and kidney-related toxicity, as well as the development of electrolyte disturbances within 3 or 7 months, was associated with shorter PFS (p= 0.011, 0.007, 0.174, and 0.02, respectively) and/or OS (p= 0.001, 0.099, 0.012, and 0.001, respectively). In multivariable Cox analysis, increasing alkaline phosphatase levels (HR = 1.93,p= 0.009), decreasing albumin (HR = 1.92,p= 0.008) and development of hyponatremia (HR = 1.79,p= 0.033) were associated with a shorter OS. The combination of static and dynamic variables significantly improved the ability of ML models to identify poor responders (shorter PFS: AUC range 0.91–0.94 vs. 0.79–0.89). Conclusions: The integration of conventional prognostic factors with the detection of significant changes in vital signs and blood tests occurring early during systemic treatment in patients with de novo mHSPC may enhance patient stratification and improve prediction of survival outcomes. Multicenter validation studies are needed to confirm these results.
背景：转移性激素敏感性前列腺癌（mHSPC）的传统预后因素通常在诊断时进行评估。然而，全身治疗期间生命体征和实验室参数的变化可能预测患者预后并提示器官特异性毒性的发生。方法：这项单中心回顾性研究纳入了2014年至2023年间接受治疗的363例初诊mHSPC患者。从治疗开始至后续七个月内，系统收集了医院数据仓库中的临床和实验室数据。采用CTCAE V5.0标准对生命体征参数和血液检测结果的变化进行分级（动态变量）。通过Cox回归分析探讨动态变量对无进展生存期（PFS）和总生存期（OS）的影响。研究构建了机器学习（ML）模型（支持向量分类器、随机森林和LGBM分类器），用于预测单一器官特异性毒性，并根据7个月PSA水平、PFS和OS识别治疗反应良好与不良的患者。比较了仅基于基线因素训练的ML模型（静态模型）与整合治疗开始后3个月和7个月内生命体征及血液检测监测变量的模型（动态模型）的性能。结果：动态模型未能提升单一器官特异性毒性的预测能力。单变量Cox分析显示，3或7个月内出现血液学、肝脏及肾脏相关毒性以及电解质紊乱，与较短的PFS（p值分别为0.011、0.007、0.174和0.02）和/或OS（p值分别为0.001、0.099、0.012和0.001）相关。多变量Cox分析表明，碱性磷酸酶水平升高（HR=1.93，p=0.009）、白蛋白降低（HR=1.92，p=0.008）以及低钠血症的发生（HR=1.79，p=0.033）与较短的OS相关。静态与动态变量的结合显著提升了ML模型识别治疗反应不良患者（PFS较短）的能力（AUC范围0.91–0.94 vs. 0.79–0.89）。结论：在初诊mHSPC患者中，将传统预后因素与全身治疗早期出现的生命体征和血液检测显著变化相结合，可优化患者分层并提升生存结局的预测能力。需开展多中心验证研究以确认这些结果。
肿瘤(癌症)患者之家