基于多模态特征融合的高血压危险分层预测与评估

Prediction and Evaluation of Hypertension Risk Based on Multimodal Feature Fusion

摘要: 机器学习应用于高血压危险分层预测，经常由于特征提取困难且异常值多、模型调优成本高，而导致模型预测精度低、泛化性能差。本文提出了一种多模态融合模型。该模型构建了多长短期记忆网络串联结构，实现无超参数的脉搏波特征提取，降低特征提取成本，提高特征区分度。使用三种不同的机器学习方法分别对脉搏波、理化、证素特征进行分类，采用基于树结构估计的贝叶斯优化算法，动态优化各机器学习模型超参数，降低超参数优化成本。采用DS证据理论冲突消解策略，减少各模型预测冲突，实现多模态特征决策级融合。本文采用上海中医药大学附属中西医结合医院体检中心等医院提供的临床数据。5-Fold交叉验证后分类模型评估指标F1-score、Accuracy、Recall、Specificity、AUC值分别为：89.1%、90.9%、89.3%、94.9%、97.7%。与典型模型相比，本文方法具有较高的分类精度和泛化性能。此外，本文分别基于最近邻、支持向量机、极限梯度提升树算法，研究了脉搏波脉图特征与高血压危险分层的相关性，深入挖掘潜在的风险因素，为高血压临床诊断提供有效参考。

Abstract: The application of machine learning in hypertension risk prediction is often hindered by difficulties in feature extraction, the presence of outliers, and high tuning costs, which result in low prediction accuracy and poor generalization performance. This paper proposes a multimodal fusion model. The model constructs a multi-long short-term memory (LSTM) network concatenation structure to achieve hyperparameter-free pulse wave feature extraction, reducing feature extraction costs and improving feature discrimination. Three different machine learning methods are used to classify pulse wave, physicochemical, and syndromic features, while a tree-structured Bayesian optimization algorithm is employed to dynamically optimize the hyperparameters of each machine learning model, thus reducing the cost of hyperparameter optimization. A DS evidence theory conflict resolution strategy is used to reduce prediction conflicts between models, enabling decision-level fusion of multimodal features. The clinical data used in this study were provided by the Health Check-up Center of the Affiliated Hospital of Shanghai University of Traditional Chinese Medicine, among others. After 5-fold cross-validation, the evaluation metrics are as follows: F1-score 89.1%, Accuracy 90.9%, Recall 89.3%, Specificity 94.9%, and AUC 97.7%. Compared with typical models, the proposed method demonstrates higher classification accuracy and generalization. In addition, based on k-nearest neighbors, support vector machines, and extreme gradient boosting tree algorithms, the paper explores the correlation between pulse wave morphology features, and hypertension risk, uncovering potential risk factors, and providing effective references for clinical hypertension diagnosis.