Frequency-Aware Masked Autoencoders for Human Activity Recognition using Accelerometers.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Wearable accelerometers are key for continuous physical activity monitoring.
• Supervised learning for human activity recognition (HAR) is limited by scarce labeled data.
• Self-supervised pretraining on large unlabeled datasets offers a promising alternative for HAR.

Purpose of the Study:

• To investigate self-supervised pretraining using a time-series transformer masked autoencoder (MAE) for HAR.
• To introduce and evaluate novel spectrogram-based loss functions: log-scale mean-magnitude (LMM) and log-scale magnitude variance (LMV).
• To compare LMM and LMV losses against mean squared error (MSE) for MAE pretraining.

Main Methods:

• Utilized a time-series transformer masked autoencoder (MAE) for self-supervised pretraining.
• Developed and tested LMM and LMV spectrogram-based loss functions.
• Pretrained models on the large unlabeled UK Biobank accelerometry dataset (n=109k).
• Evaluated downstream HAR performance using linear probing on a labeled dataset.

Main Results:

• Pretraining with LMM loss significantly improved HAR performance (12.7% increase in subject-wise F1 score) compared to MSE loss.
• LMM-pretrained transformer models outperformed a state-of-the-art ResNet-based HAR model (+9.8% F1) with linear probing.
• Adding LMV loss to LMM loss did not enhance, and slightly decreased, performance.
• LMM loss demonstrated robustness and effectiveness for MAE pretraining on accelerometer data for HAR.

Conclusions:

• The LMM loss function is effective for self-supervised pretraining of MAE models for HAR using accelerometer data.
• Self-supervised pretraining of sequence-based models holds significant potential for advancing free-living HAR.
• This approach enables more accurate physical activity monitoring, crucial for mobility assessment, rehabilitation, and chronic disease management.