The Use of a Smartphone to Assess the Two-Minute Step Test: Validity of Machine Learning Compared to Analytical Data...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning in Health

Background:

• The 2-Minute Step Test (2MST) is widely used but typically scored only by step count, neglecting movement quality.
• Assessing thigh kinematics during the 2MST could provide richer insights into physical performance and functional mobility.

Purpose of the Study:

• To evaluate the efficacy of using a smartphone sensor to quantify thigh kinematics during the 2MST.
• To compare the accuracy of machine learning (ML) versus analytical data analysis (AA) for processing smartphone signals against a motion capture (ground truth) reference.

Main Methods:

• Eighty-four healthy adults performed the 2MST while holding a smartphone to their thigh.
• Thigh angular velocity was recorded simultaneously using motion capture (Vicon) and the smartphone.
• Smartphone data underwent processing via both analytical (adaptive Butterworth filtering) and machine learning (stacked regression) approaches.

Main Results:

• Step cycles and duration were consistent across all methods.
• Machine learning analysis of smartphone data yielded peak thigh angular velocity (304 ± 37°·s^-1) with minimal bias (1.0°·s^-1) and narrow limits of agreement (-15.4-17.5) compared to motion capture.
• Analytical data analysis showed greater bias (25.5°·s^-1) and wider limits of agreement (-49.8-100.8), indicating approximately 8% error.

Conclusions:

• Smartphones held to the thigh can effectively quantify 2MST performance, including step count, timing, and thigh angular velocity.
• Machine learning data processing significantly improves the accuracy of smartphone-based kinematic analysis, closely matching gold-standard motion capture.
• This approach offers a feasible and accurate method for assessing 2MST performance beyond simple step counting.