A machine learning approach for the design optimization of a multiple magnetic and inertial sensors wearable system for the spine mobility assessment.

Background: Recently, magnetic and inertial measurement units (MIMU) based systems have been applied in the spine mobility assessment; this evaluation is essential in the clinical practice for diagnosis and treatment evaluation. The available systems are limited in the number of sensors, and neither develops a methodology for the correct placement of the sensors, seeking the relevant mobility information of the spine.

Methods: This work presents a methodology for analyzing a system consisting of sixteen MIMUs to reduce the amount of information and obtain an optimal configuration that allows distinguishing between different body postures in a movement. Four machine learning algorithms were trained and assessed using data from the range of motion in three movements (Mov.1-Anterior hip flexion; Mov.2-Lateral trunk flexion; Mov.3-Axial trunk rotation) obtained from 12 patients with Ankylosing Spondylitis.

Results: The methodology identified the optimal minimal configuration for different movements. The configuration showed good accuracy in discriminating between different body postures. Specifically, it had an accuracy of 0.963 ± 0.021 for detecting when the subject is upright or bending in Mov.1, 0.944 ± 0.038 for identifying when the subject is flexed to the left or right in Mov.2, and 0.852 ± 0.097 for recognizing when the subject is rotated to the right or left in Mov.3.

Conclusions: Our results indicate that the methodology developed results in a feasible configuration for practical clinical studies and paves the way for designing specific IMU-based assessment instruments.

Trial registration: Study approved by the Local Ethics Committee of the General Hospital of Mexico "Dr. Eduardo Liceaga" (protocol code DI/03/17/471).