Mittal Pooja, Sharma Yogesh Kumar, Rai Anjani Kumar, Simaiya Sarita, Lilhore Umesh Kumar, Kumar Vimal
Department of Computer Science & Applications, Maharshi Dayanand University, Rohtak, Haryana, India.
Department of Computer Science & Engineering, KoneruLakshmaiah Education Foundation, Green Field, Vaddeswaram, Guntur, Andhra Pradesh, India.
Sci Rep. 2025 Jan 2;15(1):328. doi: 10.1038/s41598-024-83357-9.
A dual-stage model for classifying Parkinson's disease severity, through a detailed analysis of Gait signals using force sensors and machine learning approaches, is proposed in this study. Parkinson's disease is the primary neurodegenerative disorder that results in a gradual reduction in motor function. Early detection and monitoring of the disease progression is highly challenging due to the gradual progression of symptoms and the inadequacy of conventional methods in identifying subtle changes in mobility. The proposed dual-stage model utilized a hypertuned Random Forest Tree (RFT) to classify the subjects into PD and non-PD classes at Stage 1 and a hypertuned Ensemble Regressor (ER) to predict the severity of illness at Stage 2. Further, we have implemented the proposed model on the data signals gathered from both feet of 166 participants using Vertical Ground Reaction Force Sensors (VGRF). The dataset comprised 93 persons with Parkinson's disease and 73 healthy controls. The dataset (imbalance) collected from both feet is passed to the preprocessing phase (for balancing data using the SMOTE method), followed by the feature extraction phase to extract features related to time, frequency, spatial, and temporal features domains that are highly effective for detecting and assigning severity levels of PD. A Recursive Feature Elimination method is also used to select the optimal set of features to improve the model performance. It is acknowledged that the early detection of Parkinson's disease is contingent upon critical parameters, including stride length, stance duration, swing interval, double limb support, step time, and step length. The crucial evaluation metrics used for evaluating model performance include accuracy, mean absolute error, and root mean square error. The findings indicate that the suggested model significantly surpasses current methodologies. It attained an accuracy of 97.5 ± 2.1%, Sensitivity of 97% ± 2.5%, and average Specificity of 95% ± 2.2% in differentiating between PD and non-PD participants utilizing RFT and evaluated disease severity with an average accuracy of 96.4 ± 2.3%, an average mean absolute error of 0.065 ± 0.024, and a root mean square error of 0.080 ± 0.06. The results indicate that the proposed dual-stage model is exceptionally successful in the early detection and severity assessment of Parkinson's disease and demonstrates better efficacy than alternative models.
本研究提出了一种用于帕金森病严重程度分类的双阶段模型，该模型通过使用力传感器和机器学习方法对步态信号进行详细分析。帕金森病是导致运动功能逐渐减退的主要神经退行性疾病。由于症状的逐渐发展以及传统方法在识别运动细微变化方面的不足，对该疾病进展的早期检测和监测极具挑战性。所提出的双阶段模型在第一阶段利用超参数调整的随机森林树（RFT）将受试者分为帕金森病组和非帕金森病组，在第二阶段利用超参数调整的集成回归器（ER）预测疾病严重程度。此外，我们使用垂直地面反作用力传感器（VGRF）在166名参与者双脚采集的数据信号上实现了所提出的模型。该数据集包括93名帕金森病患者和73名健康对照者。从双脚收集的数据集（不平衡）进入预处理阶段（使用SMOTE方法平衡数据），随后进入特征提取阶段，以提取与时间、频率、空间和时间特征域相关的特征，这些特征对于检测和确定帕金森病的严重程度非常有效。还使用递归特征消除方法选择最优特征集以提高模型性能。公认帕金森病的早期检测取决于关键参数，包括步长、站立持续时间、摆动间隔、双支撑期、步时和步幅。用于评估模型性能的关键评估指标包括准确率、平均绝对误差和均方根误差。研究结果表明，所建议的模型显著优于当前方法。在使用RFT区分帕金森病参与者和非帕金森病参与者时，其准确率达到97.5 ± 2.1%，灵敏度为97% ± 2.5%，平均特异性为95% ± 2.2%，评估疾病严重程度时平均准确率为96.4 ± 2.3%，平均绝对误差为0.065 ± 0.024，均方根误差为0.080 ± 0.06。结果表明，所提出的双阶段模型在帕金森病的早期检测和严重程度评估方面非常成功，并且比其他模型具有更好的效果。
