A Masked Multi-Task Learning Approach for Otago Micro Labels Recognition.

Area of Science:

• Gerontology
• Biomedical Engineering
• Machine Learning

Background:

• The Otago Exercise Program (OEP) is crucial for older adult rehabilitation, focusing on strength, balance, and fall prevention.
• Current Human Activity Recognition (HAR) systems primarily track long exercise durations (macro activities), not individual repetitions (micro activities) within programs like OEP.

Purpose of the Study:

• To develop a novel multi-task machine learning approach for recognizing micro activities within the Otago Exercise Program.
• To address the challenge of limited dataset sizes in HAR for rehabilitation exercises.

Main Methods:

• Utilized a Transformer encoder for feature extraction and a Temporal Convolutional Network (TCN) for classification.
• Implemented masked self-supervised learning with the Transformer encoder for signal reconstruction to enhance supervised learning performance.

Main Results:

• The combined unsupervised and supervised learning approach achieved f1-scores exceeding the clinically relevant threshold of 0.8.
• Successfully identified two key clinical outcomes: counting exercise repetitions and calculating chair rising velocity.

Conclusions:

• The proposed multi-task learning model effectively recognizes micro activities in the OEP, outperforming existing HAR systems.
• This technology enables automatic monitoring of exercise intensity and difficulty, supporting personalized rehabilitation for older adults.