Category: Technology
Objective: Determine if machine learning models can classify healthy individuals and people with movement disorders using smartphone data collected during a simple balance task.
Background: The ubiquity of wearable devices has the potential to revolutionize the diagnosis, monitoring, and treatment of movement disorders. However, methods that minimize burden while also maximizing precision and accuracy of outcomes need to be developed for wider use and acceptability.
Method: We collected data from healthy controls, patients with PD, and patients with ET. Participants performed a 30s quiet standing task with feet together and a 30s tandem standing task while holding a smartphone. The time-series was segmented into one-dimensional point-to-point sub-movements based on velocity zero-crossings. These sub-movements were then normalized, and several features were extracted and used as inputs into several machine learning classifiers with a 4-fold cross-validation approach (e.g., ExtraTrees, GradientBoosting, Random Forest, KNN, SVM, AdaBoost, DecisionTree). The performance of the different models was then compared in their ability to classify healthy controls from the pooled group of patients with PD and ET, and then to classify PD from ET.
Results: Our results show that features of sub-movements extracted from smartphone data collected during balance tasks can accurately classify healthy controls from people with movement disorders. The best models provided 90% accuracy using the data from the balance task with feet together (other models ranged from 80-85%), and slightly better 95% accuracy during the tandem task (other models ranged from 75-90%). Our results also show the ability of this approach to differentiate PD from ET. The best model provided 88% accuracy using data from the balance task with feet together (other models ranged from 50-75%), and 100% accuracy using data from the tandem task (other models ranged from 0-83%). The most important features for the best performing models included a combination of frequency-based measures and nonlinear metrics.
Conclusion: This work shows that using novel data analysis approaches enables the accurate classification of healthy individuals and patients with PD and ET from data collected using a smartphone during simple 30s balance tasks. This could lead to easier large scale screening and long-term monitoring of these disorders.
To cite this abstract in AMA style:
S. Macneille, A. Kostic, A. Chu, J. Daneault. Balancing Act: Smartphone Data and Machine Learning in Movement Disorder Classification [abstract]. Mov Disord. 2024; 39 (suppl 1). https://www.mdsabstracts.org/abstract/balancing-act-smartphone-data-and-machine-learning-in-movement-disorder-classification/. Accessed October 4, 2024.« Back to 2024 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/balancing-act-smartphone-data-and-machine-learning-in-movement-disorder-classification/