CLASSIFICATION OF PATIENTS WITH CHRONIC LOW BACK PAIN AND HIGH OR LOW CENTRAL SENSITIZATION BY GAIT OUTCOMES USING MACHINE LEARNING METHODS

A major issue for interventions for patients with chronic low back pain (CLBP) is the heterogeneity of the patient population. Many studies have shown in a controlled laboratory setting that gait of patients with CLBP is different in terms of less variability in trunk rotations [1], lower velocity in preferred walking speed [2], shorter stride length [3].
One of the explanations for the inconsistent finding could be the presence of central sensitization [4] since movements may be changed during pain. Also, results might be different when assessed in an uncontrolled environment compared to walking during daily life activities.
Therefore, this study aimed to examine if patients with CLBP with high or low probability of CS (a CS Inventory score of 40-100 (CLBP+), or lower than 40 (CLBP-)) could be classified based on gait performance outcomes, obtained from gait cycles during daily life activities.
Forty-three patients with CLBP were included (24 CLBP- and 19 CLBP+). Patients wore a 3D accelerometer for about one week. From each patient, 4 days of accelerometer-data was selected randomly. For each day data, continuous gait cycles (628 for CLBP-, 571 for CLBP+) were extracted using a zero-cross method. For all gait cycles in one day, 36 gait outcomes were calculated, representing variables related pace, regularity, synchronization, smoothness, local stability, and predictability of gait. A Random Forest classifier was trained to classify CLBP- and CLBP+ groups based on gait outcomes and SHapley Additive exPlanations (SHAP) method was used to explain the differences between groups in gait outcomes.
The low and high CS groups were classified with a Random Forest method with the F1-score of 0.82, an accuracy of 81%. Eight gait outcomes were identified by SHAP as the most important in classifying 2 groups. They were index of harmonicity-V and harmonic ratio-ML (smoothness), and sample entropy-AP (predictability), and maximal-Lyapunov exponent- V/ML (stability), and stride frequency variability-ML/AP (pace), and walking regularity-ML (regularity).
The accurate classification results indicate that patients with CLBP and with high or low CS walked differently. The SHAP method shows that patients with high CS level exhibited lower smoothness, lower local stability, and less predictable in gait.