骨肉瘤患者2年肺转移风险机器学习预测模型的构建

Construction of a Machine Learning Prediction Model for the 2-Year Lung Metastasis Risk in Patients with Osteosarcoma

摘要: 目的：构建基于机器学习技术的骨肉瘤患者2年肺转移风险预测模型，并精准识别影响肺转移发生的关键预测指标，为临床早期干预提供参考依据。方法：回顾性收集青岛大学附属医院2016年1月至2025年12月期间收治的骨肉瘤患者临床资料。肺转移的定义为随访2年内通过影像学检查明确证实存在肺部转移病灶。采用Boruta算法(实施10折交叉验证，筛选被10折中超过9折判定为重要的特征)、相关性分析及LASSO回归三种方法联合筛选核心预测变量，构建8种不同的机器学习模型。结果：经多步特征筛选后，最终保留6个关键预测特征，分别为白蛋白、软组织受累、血红蛋白、碱性磷酸酶、C反应蛋白及乳酸脱氢酶。在构建的8种机器学习模型中，逻辑回归(LR)模型表现最优，其训练集曲线下面积(AUC)为0.93，测试集AUC为0.92，敏感性达92%，特异性为82%，展现出良好的泛化能力和预测可靠性。结论：基于多层感知机(MLP)构建的可解释性模型，能够有效识别骨肉瘤肺转移的高危患者，该模型预测效能稳定，具备较高的临床推广应用价值。

Abstract: Objective: To develop a machine learning-based model for predicting the 2-year risk of pulmonary metastasis in patients with osteosarcoma, and to accurately identify key predictive indicators associated with pulmonary metastasis, thereby providing evidence for early clinical intervention. Methods: Clinical data of patients with osteosarcoma treated at the Affiliated Hospital of Qingdao University from January 2016 to December 2025 were retrospectively collected. Pulmonary metastasis was defined as the presence of lung metastatic lesions confirmed by imaging examinations within a 2-year follow-up period. Core predictive variables were jointly screened using the Boruta algorithm (with 10-fold cross-validation, selecting features identified as important in more than 9 of the 10 folds), correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Subsequently, eight different machine learning models were constructed. Results: After multistep feature selection, six key predictive features were ultimately retained: albumin, soft tissue involvement, hemoglobin, alkaline phosphatase, C-reactive protein, and lactate dehydrogenase. Among the eight constructed machine learning models, the logistic regression (LR) model demonstrated the best performance, with an area under the curve (AUC) of 0.93 in the training set and 0.92 in the test set. The model achieved a sensitivity of 92% and a specificity of 82%, indicating good generalization ability and reliable predictive performance. Conclusion: An interpretable model based on a multilayer perceptron (MLP) can effectively identify osteosarcoma patients at high risk of pulmonary metastasis. The model demonstrates stable predictive performance and has considerable potential for clinical application.