According to a recent study presented at the 2023 International Congress of Parkinson’s Disease and Movement Disorders, held August 27-31, in Copenhagen, Denmark, results showed that a multilayer perceptron (MLP) machine learning approach performed the best in predicting gait dysfunction among patients with Parkinson disease (PD).1 The findings suggest that the use of such machine learning approach from diffusion tensor imaging multispectral diffusion weighted imaging can effectively predict gait dysfunction in PD.
Among 43 patients with PD, the MLP, which included 5 hidden layers and Rectified Linear Unit activation, performed the best on the classification of gait dysfunction (area under the curve [AUC] = 0.78) compared with loss distribution approach (LDA) (AUC = 0.72). Notably, the machine learning approach had a better performance in predicting gait dysfunction when compared with Random Forest (AUC = 0.69), Gradient Boosting Machines (GBM) (AUC= 0.67), and Long Short-Term Memory (LSTM) (AUC = 0.50), which had a training loss of 0.0117.2
“Recent advances in MR methodology allowing quantitative evaluation of biochemical changes and macro- and microstructural alterations as well as analytic approaches including voxel-based analyses, machine-learning techniques and other post-processing algorithms have gained growing popularity in medical image analysis offering insights into the pathophysiology underlying key symptoms such as gait dysfunction in PD,” Klaus Seppi, MD, director of the department of neurology at the Hospital Kufstein and professor for neurology at the Medical University Innsbruck, commented on the study in a statement.1
In this retrospective study, patients underwent DTI image preprocessing to generate 7 quantitative measures from 30 regions of interest (ROIs) using diffusion tensor imaging. Thus, researchers compared the performance of various machine learning techniques, including LDA, MLP, LSTM, GBM, Random Forest, XGBoost, and the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (UPDRS)-III to develop a prediction model for gait dysfunction. Hyperparameter fine-tuning was conducted using GridSearch while age, sex, UPDRS-III scores, and the ROIs, which were trained and tested using a split of 80 out of 20, were used as features in the models. Notably, the LSTM was trained for 1000 epochs utilizing a learning rate of 0.0001.
“This study suggests that a machine learning approach of DTI analysis may have potential in predicting gait dysfunction in patients with PD if confirmed by larger confirmative studies. Moreover, future studies have to explore if this pattern changes with disease progression,” Seppi responded to the study in a statement.1 Therefore, authors noted that future research using deep learning models and multimodal patient information may help to confirm these results and improve the accuracy of these predictive models.2
In PD, gait dysfunction is a debilitating symptom of the disease and can be presented with a variety of different types of severities. Thus, diffusion tensor imaging provides an unbiased assessment of gait dysfunction and may be a noninvasive biomarker for recognizing the symptoms. Also, the use of machine learning approaches proposes a potential way to build predictive models capable of including large quantities of ROIs and clinical information features for predicting gait dysfunction in patients with PD.3