In-The-Wild Detection of Intermittent Parkinsonian Tremor: A Federated, Self-Supervised Learning Approach Using...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Parkinson's disease (PD) diagnosis relies on motor and non-motor symptoms, with tremor being a key indicator.
• Early detection of PD is crucial for effective symptom management.
• Current PD detection methods face challenges in real-world applicability.

Purpose of the Study:

• To develop and evaluate a novel approach for Parkinsonian tremor detection using smartphone accelerometer data.
• To enhance representation learning by integrating contrastive pretraining within a Multiple-Instance Learning (MIL) framework.
• To extend the method for scalable and privacy-preserving deployment using federated learning.

Main Methods:

• Utilized smartphone-captured accelerometer data collected in-the-wild for Parkinsonian tremor detection.
• Integrated contrastive pretraining with a Multiple-Instance Learning (MIL) framework to leverage unlabeled data.
• Extended the model to a federated learning framework to address scalability and privacy concerns.

Main Results:

• The pretrained MIL model demonstrated superior performance compared to the baseline.
• Performance improved with an increasing size of the unlabeled dataset.
• Federated pretraining achieved performance comparable to centralized training, while preserving privacy and handling non-IID data.

Conclusions:

• The proposed method effectively detects Parkinsonian tremor using smartphone data.
• Contrastive pretraining enhances MIL framework performance for PD detection.
• Federated learning offers a scalable and privacy-preserving solution for real-world PD monitoring.