Experimental Study of Long Short-Term Memory and Transformer Models for Fall Detection on Smartwatches

Area of Science:

• Wearable technology
• Artificial intelligence
• Biomedical engineering

Background:

• Falls are a leading cause of injury deaths globally.
• Existing AI-powered wearable fall detection systems on smartwatches struggle with real-time accuracy, missing falls and generating false positives.
• Previous research explored Long Short-Term Memory (LSTM) models for fall detection, showing promise in offline testing but not always translating to real-time performance.

Purpose of the Study:

• To compare the performance of Long Short-Term Memory (LSTM) and Transformer models for real-time fall detection on smartwatches.
• To evaluate the effectiveness of different LSTM and Transformer model variants in learning fall patterns.
• To determine the optimal model for practical, real-time fall detection applications using commodity smartwatches.

Main Methods:

• Investigated and experimented with three variants of LSTM and two variants of Transformer models.
• Trained all models using fall and activity data from three distinct datasets.
• Conducted real-time testing of the trained models using the SmartFall application on smartwatches.

Main Results:

• In offline training, the Convolutional Neural Network-LSTM (CNN-LSTM) model demonstrated superior performance across all datasets compared to Transformer models.
• Despite offline performance, real-time testing revealed limitations in translating LSTM model accuracy to practical application.
• Transformer models showed potential for effective real-time fall detection due to their ability to learn long-sequence data via self-attention.

Conclusions:

• While CNN-LSTM models excel in offline fall detection analysis, Transformer models are a more suitable choice for real-time deployment on smartwatches.
• The self-attention mechanism in Transformer models aids in capturing complex, long-term patterns crucial for accurate real-time fall detection.
• Further development of Transformer-based systems could significantly improve the reliability and practicality of wearable fall detection technology.