Automated Cognitive Health Assessment Using Partially Complete Time Series Sensor Data.

Area of Science:

• Biomedical Engineering
• Digital Health
• Machine Learning

Background:

• Behavior and health are closely interconnected.
• Continuous wearable sensor data hold potential for predicting clinical health measures.
• Data gaps in continuous sensor collection necessitate strategic imputation methods.

Purpose of the Study:

• To adapt a data generation algorithm for imputing multivariate time series data.
• To create digital behavior markers from imputed data for predicting clinical health measures.

Main Methods:

• Developed a bidirectional time series generative adversarial network (GAN) for imputing missing sensor readings.
• Imputed values based on inter-variable and temporal relationships for single points or gaps.
• Extracted digital behavior markers from complete, imputed data and mapped them to clinical measures.

Main Results:

• Validated the approach using smartwatch data from 14 participants.
• Achieved an average normalized mean absolute error of 0.0197 in reconstructing omitted data.
• Machine learning models predicting clinical measures from reconstructed data showed correlations from 0.1230 to 0.7623.

Conclusions:

• Wearable sensor data collected in natural settings can provide valuable health insights.
• The developed imputation method effectively handles data gaps in wearable sensor data.
• Digital behavior markers derived from imputed data show promise in predicting clinical health outcomes.