Introduction
Chronic low back pain (cLBP) is the most disabling condition in the US and is associated with opioid use disorder. cLBP can be complex to assess, with multiple potentially interacting contributors such as spinal pathology, altered biomechanics, and psychosocial and behavioral risk factors. This multi-dimensional nature of cLBP makes it challenging for clinicians to develop personalized treatment for achieving the best outcomes. To better capture and understand the change of individual cLBP patients’ pain experience over time, it is desired that there is an efficient and accurate approach for pain assessment.
Studies published in recent years show the potential of wearable sensors contributing to efficient pain assessment [1-5]. Wearable sensors offer the advantage of continuously monitoring various physiological and biomechanical parameters associated with cLBP. Features extracted from the wearable sensors data may be associated with pain severity and its resulting impact on a person’s function, activity levels, pain-related behaviors, and psychosocial factors.
Methods
In our study, we developed a multi-sensor inertial measurement unit (IMU)-based wearable sensor system for tracking global kinematics of the lumbar region. In the in-clinic test, four wireless IMUs were attached to participants at T1/T2, T12/L1, and L5/S1, and the right lateral thigh. Kinematic data were collected during a series of common functional tests. Lumbopelvic kinematic metrics included: 1) maximum range of motion (ROM), 2) maximum velocity, 3) maximum acceleration, and 4) percent contribution of hip movement to flexion/extension (lumbopelvic rhythm).
848 patients with cLBP were arranged into normal, mild/low, moderate/medium, and high/severe groups according to their pain-related measures via four standard metrics: Pain Intensity [6], Oswestry Disability Index (ODI) [7], PROMIS Physical Function and Pain Interference standard scores.
The extracted biomechanics features were preprocessed by going through missing data handling and scaling. Multiple feature selection approaches were used to select features for machine learning.
The obtained features were used in three supervised machine learning algorithms (K nearest neighbors, Support Vector Machine, and Random Forests). The one with the highest performance went through hyperparameter tuning. Ten-fold cross validation was used to achieve a reliable average performance of the supervised machine learning.
The performance of the supervised machine learning was evaluated with accuracy, precision, recall, and F-1 score.
Results
Random Forest achieves the best performance overall. The accuracy was 69% for pain intensity, 65% for ODI, 67% for PROMIS pain interference, and 76% for PROMIS physical function. The values for precision, recall, and F-1 score are in a similar range as accuracy. These results are good for a multi-class classification. Important features identified by the algorithms include velocity, flexion, acceleration, and ROM from activities such as lumbar axial rotation, sit-to-stand, lateral bending, and flexion/extension measured at the hip, lumbar, and thoracic areas.
Discussion
In conclusion, the IMU-based wearable sensor data may be interpreted using machine learning (ML) algorithms to contribute to efficient pain assessment. IMU-based wearable sensor data, along with other assessments, may lead to the development of patient-specific models for cLBP personalized interventions, optimized treatment responses and improved outcomes.