Related Experiment Video
Updated: Jun 28, 2026
10:28
Dynamic Digital Biomarkers of Motor and Cognitive Function in Parkinson's Disease
Published on: July 24, 2019
Remote Assessment of Parkinson Disease Using Deep Learning on Structured Mouse-Trace Data From Suspected Cases:
Md Rahat Shahriar Zawad1, Zerin Nasrin Tumpa2, Lydia Sollis2
1Department of Electrical and Computer Engineering, Rice University, Houston, TX, United States.
JMIR Formative Research
|June 26, 2026
Summary
This study shows artificial intelligence models can detect Parkinson disease (PD) using mouse-tracing data from community samples, even with less specific diagnoses. This approach supports PD screening in low-resource settings where specialist diagnoses are challenging.
Area of Science:
- Artificial Intelligence
- Neurodegenerative Diseases
- Biomedical Engineering
Background:
- Parkinson disease (PD) is a global neurodegenerative disorder primarily affecting motor function.
- Current AI models for PD detection often rely on data from well-resourced settings with confirmed diagnoses.
- Specialist-confirmed labels for PD are frequently unfeasible in low-resource environments.
Purpose of the Study:
- To assess the feasibility of training AI models using data from community-recruited participants with suspected PD.
- To determine if weaker diagnostic labels, more accessible in global health settings, can yield diagnostically useful predictive signals.
- To develop a web platform for collecting structured mouse-tracing data for PD assessment.
Main Methods:
- Recruited 261 participants (73 self-reported PD, 155 non-PD, 33 suspected PD) from community organizations.
- Collected mouse-tracing data (cursor position, screen dimensions) during straight line, sine wave, and spiral wave tasks via a web platform.
- Engineered features and generated mouse trace images; developed classifiers including feed-forward neural networks, deep learning computer vision models, and multimodal models.
Main Results:
- Multimodal Vision Transformer achieved an F1 score of 0.7619 in the primary experiment (suspected PD vs. non-PD, tested on self-reported PD vs. non-PD).
- Multimodal ResNet-50 achieved an F1 score of 0.9353 in the secondary analysis (self-reported PD vs. non-PD, tested on suspected PD vs. non-PD).
- Models trained on suspected PD demonstrated meaningful performance in predicting self-reported PD, indicating feasibility of using lower-specificity labels.
Conclusions:
- Remotely collected mouse-tracing data can support PD screening AI models, even with low diagnostic specificity.
- AI models trained on suspected PD from community samples may identify signals transferable to predicting actual PD.
- Future research could involve pretraining models with weaker labels and fine-tuning with stronger clinical data.
