基于动态权重网络的自适应集成模型及疾病预测应用

Adaptive Integrated Model Based on Dynamic Weighting Network and Its Application to Disease Prediction

摘要

针对传统集成预测模型存在的固态权重局限，难以实现个性化预测的问题，构建了基于动态权重网络的自适应集成预测模型．通过元学习策略集成了逻辑回归、支持向量机、随机森林3种机器学习算法，依据关键特征动态调整各种基模型的权重，实现集成模型的自适应调节．进一步，为了验证自适应集成模型的有效性，针对急性呼吸道传染病的重症风险预测任务进行实证验证．基于 1 485 例患者的临床数据对模型进行训练和评估，结果表明：自适应集成模型在训练集上的预测准确率为99.05%，F₁分数为0.994，曲线下面积为 0.998；在测试集中预测准确率达到 99.55%，F₁分数为0.997，曲线下面积为0.998，各项评估指标均优于其他单一基模型及其他集成预测模型．特别是自适应集成模型的特异度为0.999，较基模型提高3%，表现出更强的非重症患者识别能力．此外，通过雅可比矩阵揭示关键特征与基模型权重分配之间的动态关联，分析显示，该模型能够适应关键特征的变化，自动调整基模型的权重分布，凸显更适合当前数据特征的基模型贡献．这种动态调整机制弥补了传统集成模型静态权重的不足，实现了预测策略的个性化调整．构建的自适应集成模型能够有效整合多个基模型的优势，并根据特征变化动态调整基模型权重，提供了更为准确、个性化的疾病风险预测，具有广泛的应用前景．

Abstract

To address the static weight limitations of traditional integration prediction models，which restrict personalized predictions，an adaptive integrated model was constructed based on a dynamic weighting network. This dynamic-weighting-network-based model integrated three machine learning algorithms：logistic regression，support vector machines，and random forests，using a meta-learning strategy. This facilitated the dynamic adjustment of base model weights using key features，thereby enabling the adaptive regulation of the integration model. The adaptive integrated model was empirically tested to validate its effectiveness in predicting severe risks for infectious acute respiratory diseases. Further，the adaptive integrated model was trained and evaluated using the clinical data of 1 485 patients，and the results revealed that it achieved accuracy，F₁ and area under the receiver operating characteristic curve(AUC)scores of 99.05%，0.994，and 0.998，respectively，on the training set. On the test set，it realized accuracy，F₁ and AUC scores of 99.55%，0.997，and 0.998，respectively，outperforming other individual base models and integrated prediction models in all the evaluation metrics. Notably，it realized a specificity of 0.999，representing a 3% increase over other base models，thereby demonstrating its stronger ability to identify noncritical patients. Furthermore，Jacobian matrix-based analysis revealed a dynamic link between key features and base model weight allocation. The model automatically adapts to changes in critical features，adjusting the weight distributions to highlight the most relevant base model for specific data patterns. This dynamic mechanism addresses the static weight limitations of traditional integrations，enabling the adjustments of personalized prediction strategies. Further，the adaptive integrated model effectively combines the strengths of multiple base models. By dynamically adjusting weights according to feature changes，it enhances prediction accuracy and flexibility，thereby overcoming the static weight limitations in traditional integration models. Overall，the model effectively provides a highly accurate and personalized tool for disease risk prediction，exhibiting broad application prospects.

关键词

动态权重 / 元学习 / 自适应集成模型 / 机器学习 / 疾病预测

Key words

dynamic weighting / meta-learning / adaptive integrated model / machine learning / disease prediction

参考文献

[1] Nwanosike E M, Conway B R, Merchant H A, et al. Potential applications and performance of machine learning techniques and algorithms in clinical practice：A systematic review[J]. International Journal of Medical Informatics, 2022(3):159.

[2] Bautista D, Arana E, Martı́—Bonmatı́ L, et al. Validation of logistic regression models in small samples：Application to calvarial lesions diagnosis[J]. Journal of Clinical Epidemiology, 1999, 52(3):237-241.

[3] Hadzikadic M, Hakenewerth A, Bohren B, et al. Concept formation vs logistic regression：Predicting death in trauma patients[J]. Artificial Intelligence in Medicine, 1996, 8(5):493-504.

[4] Campos—Filho N, Franco E L F. Microcomputer—assisted multivariate survival data analysis using Cox’s proportional hazards regression model[J]. Computer Methods and Programs in Biomedicine, 1990, 31(2):81-87.

[5] Schelling M, Braun M, Kuhn W, et al. Combined transvaginal B—mode and color doppler sonography for differential diagnosis of ovarian tumors：Results of a multivariate logistic regression analysis[J]. Gynecologic Oncology, 2000, 77(1):78-86.

[6] Ohno—Machado L. A comparison of Cox proportional hazards and artificial neural network models for medical prognosis[J]. Computers in Biology and Medicine, 1997, 27(1):55-65.

[7] Bramer M. Avoiding Overfitting of Decision Trees[M]. London：Springer London, 2013.

[8] Lim W S, van der Eerden M M, Laing R, et al. Defining community acquired pneumonia severity on presentation to hospital：An international derivation and validation study[J]. Thorax, 2003, 58(5):377-382.

[9] Kannel W B, McGee D, Gordon T. A general cardiovascular risk profile：The Framingham study[J]. American Journal of Cardiology, 1976, 38(1):46-51.

[10] Lebedev A V, Westman E, van Westen G J P, et al. Random forest integrations for detection and prediction of Alzheimer’s disease with a good between—cohort robustness[J]. NeuroImage：Clinical, 2014, 6:115-125.

[11] Nilsson J, Ohlsson M, Thulin L, et al. Risk factor identification and mortality prediction in cardiac surgery using artificial neural networks[J]. The Journal of Thoracic and Cardiovascular Surgery, 2006, 132(1):12-19.

[12] Majid A, Ali S, Iqbal M, et al. Prediction of human breast and colon cancers from imbalanced data using nearest neighbor and support vector machines[J]. Computer Methods and Programs in Biomedicine, 2014, 113(3):792-808.

[13] Weng S F, Jenna R, Joe K, et al. Can machine—learning improve cardiovascular risk prediction using routine clinical data?[J]. PLOS One, 2017, 12(4):e0174944.

[14] Gulshan V, Peng L, Coram M, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs[J]. JAMA, 2016, 316(22):2402-2410.

[15] Cui S Z, Wang D J, Wang Y Z, et al. An improved support vector machine—based diabetic readmission prediction[J]. Computer Methods and Programs in Biomedicine, 2018, 166:123-135.

[16] Pu C, Yu H, Roche C, et al. Random forests：A novel approach in lung cancer risk prediction modeling through machine learning[J]. Chest, 2019, 156(4):A1728.

[17] Macaulay B O, Aribisala B S, Akande S A, et al. Breast cancer risk prediction in African women using random forest classifier[J]. Cancer Treatment and Research Communications, 2021, 28:100396.

[18] Premeaux T A, Bowler S, Friday C M, et al. Machine learning models based on fluid immunoproteins that predict non—AIDS adverse events in people with HIV[J]. iScience, 2024, 27(6):109945.

[19] Guo C H, Liu M C, Lu M L. A dynamic ensemble learning algorithm based on K—means for ICU mortality prediction[J]. Applied Soft Computing, 2021, 103:107166.

[20] Zhang H S, Wang Y, Xie Y, et al. Prediction models based on machine learning algorithms for COVID—19 severity risk[J]. BMC Public Health, 2025, 25(1):1748.

[21] Finn C, Abbeel P, Levine S. Model—agnostic meta—learning for fast adaptation of deep networks[EB/OL]. https://arxiv.org/abs/1703.034000, 2017—07—18.

[22] Kingma D P, Ba J L. Adam：A method for stochastic optimization[C]// International Conference on Learning Representations(ICLR). San Diego，USA, 2015：6628106.

[23] Rumelhart D E, Hinton G E, Williams R J. Learning representations by back propagating errors[J]. Nature, 1986, 323(6088):533-536.

[24] Resendis—Antonio O. Jacobian matrix[J]. Encyclopedia of Systems Biology, 2013, 46(10):1061-1062.

基金资助

国家自然科学基金资助项目(72174138)

国家自然科学基金资助项目(11972252)