Using mobile phones for activity recognition in Parkinson's patients

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Mobile phones with accelerometers offer a promising method for objective quantification of everyday movements.
• Activity recognition aims to classify these movements into distinct categories like walking, standing, or sitting.

Purpose of the Study:

• To assess the accuracy of machine learning algorithms in classifying daily activities using smartphone accelerometer data.
• To investigate the performance differences in activity recognition between healthy subjects and Parkinson's disease patients.

Main Methods:

• Utilized accelerometer data from 18 healthy subjects and 8 Parkinson's disease patients.
• Employed standard machine learning classifiers (SVM, regularized logistic regression) for time series analysis.
• Performed cross-validation across all samples and subject-wise cross-validation for population-specific accuracy.

Main Results:

• High accuracy (96.1% healthy, 92.2% Parkinson's) achieved when training and testing within each group.
• Significant performance drop (60.3% accuracy) when applying healthy subject models to Parkinson's patients.
• Subject-wise cross-validation yielded 86.0% for healthy subjects and 75.1% for Parkinson's patients.

Conclusions:

• Activity recognition algorithms trained on healthy populations are not reliable for individuals with motor disabilities like Parkinson's disease.
• Movement characteristics differ significantly between healthy individuals and Parkinson's patients, impacting algorithm performance.
• Tailored algorithms or population-specific training are crucial for accurate activity recognition in diverse patient groups.