Deep Learning to Predict Falls in Older Adults Based on Daily-Life Trunk Accelerometry

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• Early fall risk detection is crucial for preventing falls in older adults.
• Wearable sensors, particularly accelerometers, offer insights into daily activities and fall risk.
• Current fall risk assessment relies on biomechanical features from accelerometer data.

Purpose of the Study:

• To investigate the efficacy of deep learning models in automatically deriving fall risk features from raw accelerometer data.
• To compare the performance of deep learning architectures (CNN, LSTM, ConvLSTM) against a traditional biomechanical feature-based model.
• To evaluate the impact of multi-task learning and data preprocessing on fall risk assessment accuracy.

Main Methods:

• Utilized an existing dataset of 296 older adults.
• Implemented and compared three deep learning models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and ConvLSTM.
• Assessed models against a baseline using traditional biomechanical features.
• Investigated single-task and multi-task learning (with gender and age as auxiliary tasks).
• Evaluated the effect of data preprocessing on model performance.

Main Results:

• Deep learning models excelled at subject identity recognition but showed only marginal improvement over the baseline for fall risk assessment in single-task mode.
• Multi-task learning, incorporating gender and age, enhanced the performance of deep learning models for fall risk assessment.
• Data preprocessing led to the best performance, achieving an Area Under the Curve (AUC) of 0.75.

Conclusions:

• Deep learning models, especially when employing multi-task learning, demonstrate effectiveness in assessing fall risk using wearable sensor data.
• Multi-task learning offers a promising approach to improve fall risk prediction accuracy.
• Further research into deep learning applications for wearable sensor data analysis in gerontology is warranted.