Personalizing Activity Recognition Models Through Quantifying Different Types of Uncertainty Using Wearable Sensors.

Area of Science:

• Wearable sensor technology
• Machine learning for health applications
• Human activity recognition

Background:

• Activities of daily living (ADL) recognition is crucial for mobile health and wellness.
• Wearable sensors and machine learning offer potential for ADL recognition.
• Inter-subject variability significantly degrades ADL recognition performance for new users.

Purpose of the Study:

• To develop a deep learning framework for personalized ADL recognition.
• To maximize personalization performance while minimizing user burden (input/label solicitation).

Main Methods:

• Proposed a deep learning-assisted personalization framework.
• Employed unsupervised retraining of feature extraction and supervised fine-tuning of classification layers.
• Utilized a novel active learning model based on model uncertainty.
• Designed a Bayesian deep convolutional neural network to estimate aleatoric and epistemic uncertainties.

Main Results:

• Improved ADL recognition accuracy by 25% on average for new users compared to non-personalized models.
• Achieved an average final accuracy of 89.2% for personalized ADL recognition.
• Demonstrated that distinguishing between aleatoric and epistemic uncertainties enhances active learning effectiveness.
• Showcased higher personalization accuracy with significantly reduced user burden.

Conclusions:

• The proposed deep learning framework effectively addresses inter-subject variability in ADL recognition.
• Distinguishing between uncertainty types enables more efficient active learning for personalization.
• The method offers a practical solution for personalized mobile health applications using wearable sensors.