Benchmarking open-source step counting algorithms for wrist-worn devices.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Human Movement Analysis

Background:

• Smartwatches and fitness trackers are common for monitoring physical activity, with step count being a key metric.
• Reliable and transparent step-counting algorithms are essential for accurate health outcome assessments.

Purpose of the Study:

• To evaluate and compare the accuracy of seven open-source step-counting algorithms for wrist-worn devices.
• To identify the most effective algorithm for step counting using smartwatch accelerometry data.

Main Methods:

• Reproduced seven open-source algorithms: 3 peak detectors, 3 periodicity detectors, and 1 dummy algorithm.
• Benchmarked algorithms using data from 20 participants wearing a smartwatch and a foot-mounted sensor.
• Activities included resting, low-intensity movement, indoor walking, and outdoor walking with stops.

Main Results:

• The windowed autocorrelation algorithm demonstrated the highest accuracy, with a 22±30% mean absolute percentage error during walking.
• The dummy algorithm outperformed peak detection algorithms, emphasizing the importance of motion detection.
• Performance was evaluated on 30-second segments comparing smartwatch and foot-sensor data.

Conclusions:

• Windowed autocorrelation is a promising approach for accurate step counting in wrist-worn devices.
• Effective motion detection strategies are critical for improving step-counting algorithm performance.
• Further research into transparent and reliable algorithms is needed for wearable health technology.