Ielo Augusto, Dattola Serena, Bonanno Lilla, De Pasquale Paolo, Cacciola Alberto, Quartarone Angelo, De Cola Maria Cristina
IEEE Trans Neural Syst Rehabil Eng. 2025;33:2034-2043. doi: 10.1109/TNSRE.2025.3574999.
Resting tremor (RT) is one of the most common and debilitating symptoms of Parkinson's Disease (PD), characterized by involuntary rhythmic muscle contractions. The Unified Parkinson's Disease Rating Scale (UPDRS) 3.17 is a clinical assessment used to evaluate the amplitude of RTs, providing critical insights into the severity of this condition. However, it relies on subjective evaluation which may introduce intra- and inter-individual biases in tremor assessment. The present study evaluates the effectiveness of a Convolutional Neural Network (CNN) model for the multiclass classification of RT amplitude in PD patients and compares its performance with traditional machine learning models. A publicly available dataset containing data from 3-axis accelerometers placed on arms of 13 PD patients and 11 healthy subjects over approximately two days, including in-clinic and daily living activities (ADLs), was used. Resting data recorded during the UPDRS assessment were extracted and used to identify additional resting periods within the recordings through an automatic segmentation algorithm. At the end, for each of the selected arms, 90,000 data points were labeled based on the respective UPDRS 3.17 scores. A CNN structured into 7 layers was developed, and a 5-fold cross-validation method was employed to test the robustness of the model. Results from the best run of the most efficient combination of hyperparameters indicate that the CNN model achieves an average accuracy of 95.94% across the validation folds. The proposed model outperformed traditional machine learning techniques as Random Forest, Support Vector Machine (SVM) and Decision Trees, demonstrating superior accuracy in tremor classification. Our method exploited the ability of CNNs to classify RT amplitude efficiently, aiming at simplify diagnostic processes, enhancing the accuracy of the RT assessment in clinical settings.
静止性震颤（RT）是帕金森病（PD）最常见且使人衰弱的症状之一，其特征为不自主的有节奏肌肉收缩。统一帕金森病评定量表（UPDRS）3.17是一种用于评估静止性震颤幅度的临床评估方法，能为该病症的严重程度提供关键见解。然而，它依赖主观评估，这可能在震颤评估中引入个体内和个体间偏差。本研究评估了卷积神经网络（CNN）模型对帕金森病患者静止性震颤幅度进行多类分类的有效性，并将其性能与传统机器学习模型进行比较。使用了一个公开可用的数据集，该数据集包含来自13名帕金森病患者和11名健康受试者手臂上放置的三轴加速度计的数据，记录时间约为两天，包括门诊和日常生活活动（ADL）。提取了UPDRS评估期间记录的静止数据，并通过自动分割算法用于识别记录中的其他静止期。最后，对于每个选定的手臂，根据各自的UPDRS 3.17评分对90,000个数据点进行标记。开发了一个由七层组成的CNN，并采用五折交叉验证方法来测试模型的稳健性。超参数最有效组合的最佳运行结果表明，CNN模型在验证折上的平均准确率达到95.94%。所提出的模型优于随机森林、支持向量机（SVM）和决策树等传统机器学习技术，在震颤分类中显示出更高的准确率。我们的方法利用了CNN有效分类静止性震颤幅度的能力，旨在简化诊断过程，提高临床环境中静止性震颤评估的准确性。
