基于迭代决策树的帕金森UPDRS预测模型研究

Research on Parkinson UPDRS Prediction Model Based on GBDT

迭代决策树(GBDT)属于机器学习算法的一种,该算法具有较好的真实分布拟合能力,可用于解决大部分回归问题。根据帕金森病对不同年龄的男性和女性患者语音的影响不同这一现实依据,提出将性别和年龄这一先验知识融入到GBDT,实现对统一帕金森评定量表(UPDRS)的预测。将性别和年龄作为先验知识,对UPDRS预测模型进行模型分解;根据迭代决策树的原理,对分解后的各模型运用决策树进行模型重构,并在各自残差减少的梯度方向上迭代训练新的决策树;将得到的以叶子节点作为增益的决策树作为最终的UPDRS预测模型。在远程帕金森数据集的仿真实验中,得到的total-updrs和motor-updrs平均绝对误差值分别为4. 498 0和3. 531 8,与最小二乘法相比,分别提高了52. 19%和53. 36%,与决策树相比,分别提高了52. 66%和52. 89%。实验结果表明,根据先验知识,使用性别和年龄的组合进行预测模型分解,并对分解各模型分别进行模型重构,能够有效提高UPDRS预测的准确率。

As a kind of machine learning algorithm,gradient boosting decision tree(GBDT)can be used to solve most of the regression problem due to fine fitting ability of the true distribution.Based on the fact that the effect of Parkinson’s disease on the speech of male and female patients of different ages is different,we use the prior knowledge of gender and age into GBDT to predict unified Parkinson’s disease rating scale(UPDRS).Use sex and age as a prior knowledge to decompose the prediction model of UPDRS.Applying decision tree to reconstruct each new model and new decision tree is iteratively trained in the direction of the gradient of the respective residuals.Decision tree with leaf node as the gain is the final prediction model of UPDRS.In the simulation experiments of remote Parkinson data set,the mean absolute error(MAE)of total-updrs is 4.498 0 and the motor-updrs is 3.531 8,which are 52.19%and 53.36% higher than that of least squares method(LS),and 52.66%and 52.89%higher than that of the classification and regression tree(CART).The experiment show that GBDT based on sex and gender partition can improve the accuracy of UPDRS prediction.