Multiple-Instance Learning for In-The-Wild Parkinsonian Tremor Detection.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Parkinson's Disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms like tremor and bradykinesia.
• Early diagnosis of PD is crucial for effective patient management and treatment.
• Current automated detection methods often lack real-world applicability due to reliance on laboratory settings.

Purpose of the Study:

• To develop and validate a novel method for the automated detection of tremorous episodes in Parkinson's Disease using acceleration signals.
• To address the limitations of existing methods by focusing on in-the-wild data collection and analysis.
• To leverage advanced machine learning techniques for improved accuracy and applicability.

Main Methods:

• A Multiple-Instance Learning framework was employed, treating subjects as bags of signal segments.
• A deep learning architecture was utilized, integrating feature learning and a learnable pooling stage.
• The model was trained end-to-end on a newly collected dataset of accelerometer signals from real-world environments.

Main Results:

• The proposed deep learning method demonstrated successful automatic detection of tremorous episodes associated with Parkinson's Disease.
• Validation was performed on a novel dataset of accelerometer signals collected in-the-wild.
• The results confirm the approach's effectiveness and potential for real-world clinical utility.

Conclusions:

• The developed method offers a promising, automated solution for detecting Parkinson's Disease tremors using readily available accelerometer data.
• The study highlights the feasibility of applying deep learning and Multiple-Instance Learning for analyzing in-the-wild physiological signals.
• This approach has the potential to significantly aid in the early and accessible diagnosis of Parkinson's Disease.