Deep Learning for Activity Recognition in Older People Using a Pocket-Worn Smartphone.

Area of Science:

• Gerontology
• Biomedical Engineering
• Computer Science

Background:

• Activity recognition is crucial for monitoring older adults' health and promoting active lifestyles.
• Smartphones offer a convenient platform for collecting accelerometry data for activity recognition.

Purpose of the Study:

• To develop and evaluate a deep learning-based activity recognition algorithm using smartphone accelerometry data for older individuals.
• To compare the performance of various deep learning models, including Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) variants.

Main Methods:

• Collected smartphone accelerometry data from 53 older adults (mean age 83.8 years) performing various activities.
• Trained and tested 1D CNN, multichannel CNN, CNN-LSTM, and multichannel CNN-LSTM models for activity classification.
• Classified activities into lying, sitting, standing, transition, walking, walking upstairs, and walking downstairs.

Main Results:

• The multichannel CNN-LSTM model achieved the highest overall classification accuracy of 81.1%.
• Specific accuracies included: lying (67.0%), sitting (70.7%), standing (88.4%), transitions (78.2%), walking (88.7%), walking downstairs (65.7%), and walking upstairs (68.7%).
• The multichannel CNN-LSTM model demonstrated acceptable model and time complexity.

Conclusions:

• The multichannel CNN-LSTM model is a feasible and effective tool for smartphone-based activity recognition in older populations.
• Accurate activity recognition can support interventions aimed at increasing physical activity and improving health outcomes in older adults.