Objective detection of Parkinson's disease motor states using Lasso-selected IMUs features

Machine learning (ML) models that use data captured from Inertial Measurement Units (IMUs) are widely applied in the clinical management of Parkinson’s disease (PD). However, performance and interpretability of these models can be influenced by the selection of input features, especially when dealing with multi-dimensional data. This study investigates the performance of Lasso regularization in improving performance and simplicity of supervised and unsupervised ML models using IMUs data. Data were collected using IMUs placed on the wrists and ankles of 19 patients (14 males and 5 females) with advanced PD (mean years with disease of 10 years). Participants performed different motor tests, and three movement disorder specialists rated the severity of motor states (Off and dyskinesia) on a Treatment Response Scale (TRS). Sensor data were processed, and features were reduced by Lasso regularization and used as inputs to Support Vector Machines (SVM) for classification and regression. In addition, clustering methods were employed to align the sensor data to clinical labels. Linear SVM correctly classified Off motor state from treatment-induced dyskinesia state with an accuracy of 93.9% and 93.3%, respectively. The correlation coefficient between the predicted TRS score derived by gaussian SVM and mean TRS score of the three specialists was 0.91. The clusters derived by the clustering algorithms separated well the instances when the patients were in Off and dyskinesia motor states. Using Lasso regularization as a feature selection method coupled with ML modelsyielded good predictive performance when fusing multi-sensor and -activity data from IMUs. This approach can be used as a tool to objectively assess PD motor states and improve the management of the disease by individualizing treatments.