融合人工鱼群和随机森林算法的膝关节接触力预测-《中国医学物理学杂志》

文章信息/Info

Title:: Knee joint contact force prediction using artificial fish swarm and random forest algorithm

作者:: 卢巍1; 朱业安2; 徐唯祎2; 1.江西省人民医院康复医学科，江西南昌 330006； 2.华东交通大学交通运输与物流学院，江西南昌 330013

Author(s):: LU Wei1; ZHU Ye’an2; XU Weiyi2; 1. Department of Rehabilitation Medicine, Jiangxi Provincial People’s Hospital, Nanchang 330006, China; 2. School of Transportation and Logistics, East China Jiaotong University, Nanchang 330013, China

Keywords:: Keywords: artificial fish swarm algorithm; random forest algorithm; knee replacement; contact force prediction

摘要:: 针对膝关节接触力的测量问题提出融合人工鱼群和随机森林算法的膝关节接触力预测方法。首先，运用混沌变换构造均匀分布的种群，并引入自适应视野范围策略和自适应步长策略，获得改进的人工鱼群算法。然后，将干预前所有受试者的步态参数和膝关节接触力数据划分为训练集（70%）和验证集（30%），利用随机森林算法对训练集进行训练，并使用改进的人工鱼群算法优化随机森林模型的主要参数，获得步态参数和膝关节接触力的非线性关系，利用验证集进行验证。最后，以干预后单个受试者的步态参数和膝关节接触力对预测模型进行测试。结果表明模型在验证集和测试集上都有很高的准确性，模型在验证集上的误差表明模型能准确地学习输入和输出之间的因果关系；在测试集上的误差表明训练后的模型能够准确地将这种因果关系推广到新的输入中。

Abstract:: Abstract: A knee joint contact force prediction method combining artificial fish swarm and random forest algorithm is proposed to measure the contact force of the knee joint. Firstly, chaos transformation is used to construct an uniformly distributed population, and the adaptive vision range strategy and adaptive step strategy are introduced to obtain the improved artificial fish swarm algorithm. Then the gait parameters and knee contact force data of all subjects before intervention are divided into training set (70%) and validation set (30%). The training set is trained by random forest algorithm, and the main parameters of random forest model are optimized by improved artificial fish swarm algorithm. The obtained nonlinear relationship between gait parameters and knee joint contact force is velidated by validation set. Finally, the gait parameters and knee contact force of each subject after intervention are used to test the prediction model. The results show that the model has high accuracy in both validation set and test set. The error of the model in validation set indicates that the model can accurately learn the causality between inputs and outputs; while the error of the model in test set indicates that the trained model can precisely generalize the causality to new inputs.