Machine Learning Estimation of COVID-19 Social Distance using Smartphone Sensor Data.

Area of Science:

• Computer Science
• Public Health
• Sensor Technology

Background:

• Airborne infectious diseases, including COVID-19, spread through close contact between individuals.
• Existing methods for proximity detection are limited, necessitating technological solutions.
• Smartphone sensors offer a potential platform for real-time proximity monitoring.

Purpose of the Study:

• To systematically investigate Machine Learning (ML) methods for detecting proximity using smartphone sensors.
• To classify proximity (less than 6 feet) and estimate distance.
• To identify the most effective ML algorithms and sensor data combinations.

Main Methods:

• Extracted 20 statistical features from smartphone Bluetooth, accelerometer, and gyroscope sensor data.
• Applied ML classification and regression algorithms to the extracted features.
• Investigated the impact of elliptical filtering on sensor signal data.

Main Results:

• Elliptical filtering improved ML regression performance for accelerometer and gyroscope data.
• Key predictive features included accelerometer z-axis mean/fourth momentum, gyroscope z-axis mean/y-axis mean, and Bluetooth advertiser time/mean RSSI.
• Ensemble methods and regression trees performed best using combined sensor data.
• Proximity (< 6ft) was classified with 100% accuracy using accelerometer data and 62%-97% accuracy using Bluetooth data.

Conclusions:

• ML algorithms, particularly ensemble and regression trees, effectively detect proximity using smartphone sensor data.
• Accelerometer data provides highly accurate proximity classification.
• This technology-assisted approach can aid in mitigating the spread of airborne infectious diseases.