Imputing Missing Data In Large-Scale Multivariate Biomedical Wearable Recordings Using Bidirectional Recurrent Neural...

Area of Science:

• Biomedical Engineering
• Data Science
• Human-Computer Interaction

Background:

• Wearable sensors enable naturalistic human behavior studies in healthcare and wellness.
• Data loss from wearable sensors (e.g., movement, displacement) impedes advanced analytics like pattern recognition.
• Traditional imputation methods fail to capture temporal dynamics in multivariate time series data.

Purpose of the Study:

• To investigate a novel data imputation method for missing data in wearable sensor recordings.
• To address the limitations of conventional imputation techniques in handling temporal variations.
• To enhance the accuracy and reliability of data analytics for wearable bio-behavioral signals.

Main Methods:

• Utilized bidirectional recurrent neural networks (RNNs) with temporal activation regularization for data imputation.
• Developed a method capable of directly learning and filling in missing data points.
• Evaluated the approach on a large-scale, multimodal dataset of bio-behavioral signals.

Main Results:

• The proposed RNN-based imputation method demonstrated superior performance compared to conventional strategies.
• Experimental results confirmed the method's effectiveness in imputing missing data in multimodal time series.
• Achieved higher imputation accuracy on a large dataset collected from over 100 hospital staff.

Conclusions:

• Bidirectional RNNs with temporal activation regularization offer a robust solution for missing data in wearable sensor streams.
• This advanced imputation technique improves the accuracy of bio-behavioral signal analysis.
• The findings support the use of deep learning for reliable data handling in wearable technology applications.