【摘要】 目的 探究导致肺癌发生的潜在风险因子,并运用机器学习(ML)算法建立肺癌发病风险预测模型,为肺癌的防治提供科学依据,以实现精准预防。方法 选取广东某医院2013年1月—2023年12月收治的首次病理确诊肺癌患者共1 066例作为肺癌组,同时,按1∶1比例选取同期健康体检中心体检结果排除肺癌和(或)其他癌症人群1 066例作为对照组。收集2组研究对象的相关资料,包括一般资料、临床资料等。数据集按7∶3的比例随机分为训练集和测试集,训练集中采用最小绝对收缩和选择算子(LASSO)算法筛选有意义的特征变量纳入模型中,以逻辑回归(LR)、随机森林(RF)、多层感知器(MLP)、支持向量机(SVM)、K最近邻(KNN)、轻量梯度提升(Light GBM)、极端梯度提升(XGBoost)、决策树(DT)等8种常用算法建立ML模型,对每一个模型进行贝叶斯优化,并采用五折交叉验证方法对其进行测试,从而得到最优超参数模型的参数配置。通过绘制受试者工作特征(ROC)曲线和决策曲线分析法(DCA)曲线来评估这些模型的区分能力和临床应用的实用性。使用SHAP算法对最佳ML模型进行解释,提升模型重要特征的直观性。最后基于Stacking集成学习组合不同模型以进一步提升模型性能。结果 采用LASSO算法筛选出重要变量,包括年龄、吸烟情况、身体质量指数(BMI)、蔬果食用量、运动时间、文化程度、糖尿病史、癌症家族史、肺癌家族史、慢性支气管炎史、肺气肿史、慢性阻塞性肺病史、长期接触煤烟、长期接触烹饪油烟、长期接触柴草烟雾。在预测模型中,表现最佳的为Light GBM模型,其曲线下面积(AUC)=0.961 8、准确率=0.893 8、精确率=0.894 0、灵敏度=0.880 9、特异度=0.906 5、召回率=0.893 8、F1分数=0.893 7,综合预测性能较好,且在DCA曲线中,Light GBM模型的净收益超过了其他模型。SHAP算法分析提示BMI、年龄、肺气肿史对肺癌发病风险贡献最大。经过Stacking集成学习的新模型准确率大部分高于单一模型,其中以SVM模型作为元学习器时准确率最高,达到89.06%,其AUC为0.915 1。结论 本研究成功构建了基于Light GBM和Stacking集成学习的肺癌发病风险预测模型,确认了多项影响肺癌发病风险的预测因子。在临床实践中,通过预测因子能快速识别高风险患者,为肺癌的早期识别和预防提供了有效工具,有助于实施更加有效的肺癌防治措施。更多还原

【Abstract】 Objective To explore the potential risk factors for the occurrence of lung cancer and use machine learning(ML) algorithms to establish a lung cancer risk prediction model, providing scientific evidence for the prevention and treatment of lung cancer and achieving precise prevention. Methods A total of 1,066 newly diagnosed lung cancer patients, confirmed by pathology, admitted to a hospital in Guangdong from January 2013 to December 2023, were selected as the lung cancer group. Simultaneously, 1,066 individuals with no lung cancer or other cancers, selected from a health check-up center at a 1∶1 ratio, were included as the control group. Data, including demographic information and clinical data, were collected from both groups. The dataset was randomly divided into a training set(70%) and a testing set(30%). The most meaningful features were selected for the model using the Least Absolute Shrinkage and Selection Operator(LASSO) algorithm. Eight common ML algorithms were employed to establish prediction models: Logistic Regression(LR), Random Forest(RF), Multilayer Perceptron(MLP), Support Vector Machine(SVM), K-Nearest Neighbors(KNN), Light Gradient Boosting Machine(Light GBM), Extreme Gradient Boosting(XGBoost), and Decision Tree(DT). Bayesian optimization was applied to each model, and five-fold cross-validation was performed to test the best hyperparameters for each model. The performance of these models was evaluated using Receiver Operating Characteristic(ROC) curves and Decision Curve Analysis(DCA). The SHAP algorithm was used to interpret the best-performing model, enhancing the visualization of important features. A new model was built based on Stacking ensemble learning to further promote model performance. Results The LASSO algorithm identified important variables, including age, smoking status, body mass index(BMI), fruit and vegetable consumption, exercise duration, education level, diabetes history, family history of cancer, family history of lung cancer, chronic bronchitis, emphysema history, chronic obstructive pulmonary disease(COPD) history, long-term exposure to coal smoke, long-term exposure to cooking oil smoke, and long-term exposure to smoke from firewood. In the prediction model,the best performance was observed in the Light GBM model,which had an area under the curve(AUC) of 0.961 8,accuracy of 0.893 8,precision of 0.894 0,sensitivity of 0.880 9,specificity of 0.906 5,recall of 0.893 8,and F1 score of 0.893 7,showing good overall predictive performance. In the DCA curve, the net benefit of the Light GBM model exceeded that of the other models. SHAP analysis indicated that BMI, age, and emphysema history contributed most to lung cancer risk. The new model, based on Stacking ensemble learning, showed an accuracy higher than that most of the single models, with the SVM model as the meta-learner achieving the highest accuracy of 89.06%and an AUC of 0.915 1. Conclusion This study successfully constructed a lung cancer risk prediction model based on Light GBM and Stacking ensemble learning, confirming multiple predictive factors for lung cancer risk. In clinical practice, the use of these predictive factors can quickly identify high-risk patients, providing an effective tool for early identification and prevention and contributing to more effective prevention and treatment measures for lung cancer.更多还原