DS-MS-TCN: Otago Exercises Recognition With a Dual-Scale Multi-Stage Temporal Convolutional Network

Area of Science:

• Gerontology
• Biomedical Engineering
• Computer Science

Background:

• The Otago Exercise Program (OEP) is vital for older adult rehabilitation, focusing on balance and strength.
• Previous wearable sensor-based OEP recognition methods lacked accuracy and robustness.
• Recognizing OEP exercises in daily life is challenging but crucial for effective monitoring.

Purpose of the Study:

• To develop a robust system for recognizing OEP exercises using a single waist-mounted Inertial Measurement Unit (IMU).
• To introduce a novel Dual-Scale Multi-Stage Temporal Convolutional Network (DS-MS-TCN) for precise exercise recognition.
• To enhance Human Activity Recognition (HAR) by recognizing individual exercise repetitions.

Main Methods:

• Utilized a single waist-mounted IMU for data collection from 36 older adults in labs and 7 at home.
• Proposed a Dual-Scale Multi-Stage Temporal Convolutional Network (DS-MS-TCN) for two-level sequence-to-sequence classification.
• Implemented a single loss function for both micro (repetition) and macro (exercise) label recognition.

Main Results:

• The DS-MS-TCN model achieved f1-scores over 80% and Intersection over Union (IoU) f1-scores over 60% for all four evaluated exercises.
• The proposed model outperformed existing state-of-the-art deep learning models and previous sliding window techniques.
• Eliminated the need for post-processing and window size tuning, simplifying the recognition process.

Conclusions:

• The DS-MS-TCN model offers a significant advancement in accurately recognizing OEP exercises in older adults.
• This approach enhances the robustness and accuracy of HAR systems for rehabilitation.
• The novel perspective of recognizing individual exercise repetitions opens new avenues for HAR research.