基于步行速度和计时起立行走测试预测帕金森病

Prediction of Parkinson’s Disease Based on Gait Speed and Timed Up and Go Test

摘要: 背景：帕金森病(Parkinson’s disease, PD)是一种影响全球数百万人的生活质量的神经系统退行性疾病，其发病率随着年龄的增长而显著增加。PD患者步态障碍，步行速度和平衡功能等均异于常人。近年来，陆续开展了早期PD预测模型的研究，但基于步态改变的机器学习对预测早期PD的研究尚少。目的：使用机器学习，整合步行速度和计时起立行走测试(Timed Up and Go, TUAG)，建立一个能够早期预测PD的模型。方法：研究使用了来自Physionet上的“Gait in Parkinson’s Disease”数据集中的88名特发性帕金森病患者和72名健康对照者的相关数据，建立逻辑回归、随机森林、支持向量机、LightGBM、XGBoost和CatBoost等六种机器学习模型，并用Grid Search CV网格搜索和5折交叉验证进行参数寻优。计算最佳超参数组合下各模型在测试集上的性能并绘制ROC曲线。应用了SHapley Additive exPlanations (SHAP)框架来解释最佳模型。结果：随机森林模型具有最高的AUC (0.771, 95%CI = 0.652~0.890)，但CatBoost模型在所有评估指标上表现最佳，其准确率为0.708，精确率为0.727，召回率为0.708，F1得分为0.695，AUC为0.766 (95%CI = 0.646~0.886)。基于SHAP值的模型解释进一步揭示了步行速度和TUAG对响模型预测结果的影响。结论：基于步行速度和TUAG的机器学习模型可准确地预测早期PD，其中，CatBoost模型具有较高的性能。

Abstract: Background: Parkinson’s disease (PD) is a neurodegenerative disorder that affects the quality of life of millions of people worldwide, and its incidence increases significantly with age. PD patients have gait disturbances, and their walking speed and balance function are different from those of normal people. In recent years, studies on early PD prediction models have been carried out successively, but there are few studies on machine learning based on gait changes for predicting early PD. Objective: To use machine learning to integrate walking speed and the Timed Up and Go (TUAG) test to establish a model that can predict PD in the early stage. Methods: The relevant data of 88 patients with idiopathic Parkinson’s disease and 72 healthy controls from the “Gait in Parkinson’s Disease” dataset on Physionet were used in this study. Six machine learning models, including logistic regression, random forest, support vector machine, LightGBM, XGBoost and CatBoost, were established, and Grid Search CV grid search and 5-fold cross-validation were used for parameter optimization. The performance of each model on the test set under the best combination of hyperparameters was calculated and the ROC curve was drawn. The SHapley Additive exPlanations (SHAP) framework was applied to interpret the best model. Results: The random forest model had the highest AUC (0.771, 95%CI = 0.652~0.890), but the CatBoost model performed best in all evaluation indicators. Its accuracy was 0.708, precision was 0.727, recall was 0.708, F1 score was 0.695, and AUC was 0.766 (95%CI = 0.646~0.886). The model interpretation based on SHAP values further revealed the influence of walking speed and TUAG on the model prediction results. Conclusion: The machine learning model based on walking speed and TUAG can accurately predict early PD. Among them, the CatBoost model has relatively high performance.