Towards On-Device Dehydration Monitoring Using Machine Learning from Wearable Device's Data.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning

Background:

• Advances in wearable sensors and miniaturization enable new health monitoring applications.
• Hydration monitoring is crucial for athletes, soldiers, outdoor workers, and individuals with thirst impairment or communication difficulties.

Purpose of the Study:

• To develop and evaluate machine learning models for hydration monitoring using data from various wearable sensors.
• To predict the last drinking time and alert users when hydration levels exceed a threshold.

Main Methods:

• Utilized data from accelerometer, magnetometer, gyroscope, galvanic skin response, photoplethysmography, temperature, and barometric pressure sensors.
• Integrated sensor data with activity and personal features to train machine learning models.
• Compared Extra Trees, Random Forest, and Deep Neural Network models for predictive accuracy and on-device suitability.

Main Results:

• The Extra Trees model demonstrated the lowest prediction error on unseen data.
• Random Forest offered a balance between accuracy and reduced training time.
• Deep Neural Network models provided a small model size suitable for memory-constrained wearable devices.

Conclusions:

• Machine learning models utilizing wearable sensor data show promise for effective hydration monitoring.
• Model selection for on-device deployment requires balancing predictive accuracy, training time, and memory footprint.
• Further embedded on-device testing is necessary to validate performance and assess power consumption.