Tu Huakang, Zhao Yunfeng, Cui Jiameng, Lu Wanzhu, Sun Gege, Xu Xiaohang, Hu Qingfeng, Hu Kejia, Wu Ming, Wu Xifeng
Center of Clinical Big Data and Analytics of the Second Affiliated Hospital and School of Public Health, Zhejiang University School of Medicine, Hangzhou 310058, China.
Department of Thoracic Surgery, The Second Affiliated Hospital, Zhejiang University School of Medicine, 88 Jiefang Rd., Hangzhou 310009, China.
Cancers (Basel). 2025 May 13;17(10):1651. doi: 10.3390/cancers17101651.
Lung cancer is a leading cause of cancer-related mortality worldwide, often diagnosed in advanced stages, making early detection critical. This study aimed to evaluate the performance of various machine learning models in predicting lung cancer risk based on epidemiological questionnaires, comparing them with traditional logistic regression models.
A retrospective case-control study was conducted using data from 5421 lung cancer cases and 10,831 matched controls. The dataset included a wide range of demographic, clinical, and behavioral risk factors from epidemiological questionnaires. We developed and compared multiple machine learning algorithms, including LightGBM and stacking ensemble models, alongside logistic regression for predicting lung cancer risk. Model performance was evaluated using accuracy, area under the curve (AUC), and recall.
The stacking model outperformed traditional logistic regression, achieving an AUC of 0.887 (0.870-0.903) compared to 0.858 (0.839-0.878) for logistic regression. LightGBM also performed well, with an AUC of 0.884 (0.867-0.901). The stacking model achieved an accuracy of 81.2%, with a recall of 0.755, higher than the logistic regression model's accuracy of 79.4%. Compared to classical lung cancer prediction models (LLP and PLCO), the logistic regression and ML models improved AUC by 12% to 27%.
Integrating machine learning models into lung cancer screening programs can significantly enhance early detection efforts. Machine learning approaches, such as LightGBM and stacking, offer improved accuracy and predictive power over traditional models. However, efforts to enhance model interpretability through explainable AI techniques are necessary for broader clinical adoption.
肺癌是全球癌症相关死亡的主要原因，通常在晚期才被诊断出来，因此早期检测至关重要。本研究旨在评估各种机器学习模型基于流行病学调查问卷预测肺癌风险的性能，并将其与传统逻辑回归模型进行比较。
采用回顾性病例对照研究，使用了5421例肺癌病例和10831例匹配对照的数据。数据集包括来自流行病学调查问卷的广泛的人口统计学、临床和行为风险因素。我们开发并比较了多种机器学习算法，包括LightGBM和堆叠集成模型，以及用于预测肺癌风险的逻辑回归。使用准确率、曲线下面积（AUC）和召回率评估模型性能。
堆叠模型优于传统逻辑回归，AUC为0.887（0.870 - 0.903），而逻辑回归的AUC为0.858（0.839 - 0.878）。LightGBM也表现良好，AUC为0.884（0.867 - 0.901）。堆叠模型的准确率为81.2%，召回率为0.755，高于逻辑回归模型79.4%的准确率。与经典肺癌预测模型（LLP和PLCO）相比，逻辑回归和机器学习模型将AUC提高了12%至27%。
将机器学习模型整合到肺癌筛查项目中可以显著加强早期检测工作。诸如LightGBM和堆叠等机器学习方法比传统模型具有更高的准确率和预测能力。然而，通过可解释人工智能技术提高模型可解释性的努力对于更广泛地应用于临床是必要的。
