基于机器学习的胰腺癌患者预后生存分析模型

A Machine Learning-Based Model for Prognostic Survival Analysis in Pancreatic Cancer Patients

摘要: 目的：针对胰腺癌预后预测中传统回归模型处理高维非线性特征受限，以及深度学习缺乏临床解释性的问题，构建一种高精度且具有强解释性的生存分析模型。方法：基于591例患者数据，选取XGBoost、SVM、MLP与CNN作为基础模型。提出自适应加权与分层融合的双阶段策略进行集成优化，并利用SHAP技术对24个临床特征进行量化解释分析。结果：该模型准确率达90.0%，AUC为0.90，召回率89.0%，相比固定权重融合提升了2%。自适应加权与层次化融合分别贡献了1.5%和0.5%的效能增幅。年龄、医疗可及性及加工食品饮食是影响预后的关键因素。结论：该方法有效解决了单一模型在复杂疾病预测中的不足，在保证高准确率的同时实现了临床可解释性，为制定个性化治疗方案提供了可靠依据。

Abstract: Objective: To address the limitations of traditional regression models in handling high-dimensional non-linear features and the lack of clinical interpretability in deep learning models, this study aims to construct a survival analysis model for pancreatic cancer that achieves both high precision and strong interpretability. Methods: Based on data from 591 patients, XGBoost, Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) were selected as base models. A two-stage ensemble optimization strategy was proposed, incorporating adaptive weighting and hierarchical fusion. Furthermore, SHAP (SHapley Additive exPlanations) technology was employed to quantitatively interpret 24 clinical features. Results: The proposed model achieved an accuracy of 90.0%, an Area Under the Curve (AUC) of 0.90, and a recall rate of 89.0%, representing a 2% improvement over fixed-weight fusion methods. Specifically, adaptive weighting and hierarchical fusion contributed to performance gains of 1.5% and 0.5%, respectively. Feature analysis identified age, healthcare accessibility, and processed food intake as critical prognostic factors. Conclusion: This approach effectively overcomes the limitations of single models in complex disease prediction. By ensuring high accuracy while maintaining clinical interpretability, the model provides a reliable basis for developing personalized treatment plans.