比较基于机器学习的不同临床模型框架在非转移性胰头癌、胰体癌和胰尾癌患者总生存率和特异生存率方面的稳定性
To Compare the Stability of Different Clinical Model Frameworks in the Overall Survival Rate and Specific Survival Rate of Patients with Non-Metastatic Pancreatic Head Cancer, Pancreatic Body Cancer and Pancreatic Tail Cancer Based on Machine Learning

近年来，预测模型在临床实践中的应用越来越广泛，针对不同患者的不同临床预测模型层出不穷。机器学习在医学中的应用正在逐渐增加。因此，本文基于机器学习的视角，研究非转移性胰头癌、胰体癌和胰尾癌患者的5年和全数据总生存期(OS)和癌症特异性生存期(CSS)，并尝试探讨整合模型、线性模型和生存树模型之间的差异性和稳定性。方法：基于机器学习技术，我们围绕患者的基本和临床信息构建了模型。选择了临床回顾性医学分析建模中目前最常用的七种模型进行比较，评估了不同类型和相同类型之间连续和分类变量数据的区分能力和准确性。研究考虑了两个终点结果：5年OS和全数据CSS。利用C-指数(一致性指数)、Brier分数、校准曲线以及净再分类指数(NRI)评估了模型的性能。结果：从2000年至2018年，共收集了6019例病理学确认的胰腺头、体和尾部癌症患者的数据。经过严格筛选，最终纳入研究的病例为3675例。研究显示，模型在预测CSS方面的准确性略优于OS。值得注意的是，梯度提升生存分析(GBSA)在各种变量类型和生存期间的CSS预测中表现最佳，无论是使用连续变量(C-指数：0.753，95% CI：0.741~0.765)还是分类变量(0.743, 0.735~0.751)。NRI分析显示，相较于分类变量，对于与OS相关的连续变量，应用Cox比例风险(CoxPH)生存分析提高了5年生存的预测能力30.5%，CSS模型提高了26.8%。NRI的散点图显示了模型之间在预测能力上的差异。结论：在所研究的模型中，GBSA表现出最高的预测能力和区分度。此外，随着临床指标的细化，多变量模型的预测能力可能会进一步提升。基于机器学习的临床前模型的整合在未来可能为肿瘤患者提供更精确的个性化治疗方案。
In recent years, the integration of predictive models into clinical practice has gained momentum, revolutionizing patient care in oncology. This study employs a machine learning perspective to in-vestigate the 5-year and long-term overall survival (OS) as well as cancer-specific survival (CSS) in patients with non-metastatic pancreatic cancer. The aim is to discern disparities and reliability among integrated, linear, and survival tree models. Methods: Utilizing machine learning techniques, we constructed models using essential patient data. Seven commonly employed models in retro-spective clinical analysis were selected for comparison, evaluating their discriminative power and accuracy for continuous and categorical variables within and between different cancer types. Two outcome measures were considered: 5-year OS and full-data CSS. Model performance was assessed using the Concordance index (C-index), Brier score, calibration curve, and Net Reclassification Index (NRI). Results: From 2000 to 2018, a total of 6019 pathologically confirmed pancreatic head, body, and tail cancer patients were collected. Following rigorous screening, 3675 patients were included in the study. The models exhibited slightly superior accuracy in predicting CSS compared to OS. No-tably, Gradient Boosting Survival Analysis (GBSA) outperformed other models in predicting CSS for both continuous (C-index: 0.753, 95% CI: 0.741~0.765) and categorical variables (0.743, 0.735~0.751) across different variable types and survival periods. The NRI analysis revealed nota-ble enhancements in predictive power when employing Cox proportional hazards (CoxPH) Survival Analysis for continuous variables in both OS (30.5% improvement) and CSS (26.8% improvement) compared to categorical variables. Scatter plots of NRI highlighted variations in prediction capabil-ity