A Deep Learning Framework for the Remote Detection of Parkinson'S Disease Using Smart-Phone Sensor Data

Area of Science:

• Neurology
• Computer Science
• Biomedical Engineering

Background:

• Wearable sensors offer objective, longitudinal Parkinson's disease (PD) assessment outside clinical settings.
• Remote data collection faces challenges like data corruption, impacting analysis.
• High-frequency, remote data is crucial for accurate PD classification and severity prediction.

Purpose of the Study:

• To evaluate machine learning and deep learning algorithms for PD classification using smartphone-based Alternate Finger Tapping test data.
• To compare the performance of traditional machine learning with deep learning approaches in classifying PD without explicit feature engineering.
• To explore the potential of deep learning in analyzing remotely collected, potentially corrupted, health data.

Main Methods:

• A cohort of 1,815 participants (949 with PD, 866 controls) provided Alternate Finger Tapping test data via smartphones.
• Implemented and compared two traditional machine learning algorithms against two deep learning models.
• Assessed deep learning's ability to classify PD without requiring a predefined feature set.

Main Results:

• Deep learning approaches demonstrated capability in classifying Parkinson's disease (PD).
• Deep learning models often outperformed traditional machine learning methods in PD classification accuracy.
• Learned features in convolutional neural networks showed similarities to manually extracted features from successful classifications.

Conclusions:

• Deep learning is suitable for assessing Parkinson's disease (PD) with large, remotely collected datasets.
• The study highlights the potential of deep learning for objective, high-frequency PD monitoring.
• Addressing challenges in remote data analysis is key to leveraging wearable sensor technology for neurological disorders.