Feature Modeling for Interpretable Low Back Pain Classification Based on Surface EMG

Abstract

Low back pain (LBP) is a global health-problem phenomenon. Most patients are categorized as non-specific, thus requiring an individualized approach which still poses a major challenge. In this paper, sEMG recordings from two pairs of lumbar muscle sites were collected during an isometric trunk extension exercise. Ninety-one subjects were included in the study; 29 patients with non-specific chronic LBP (CLBP), 25 patients with radiculopathy (RLBP), and 37 control healthy subjects (HS). Six best-performing timedomain raw features were employed to model contextual secondary feature groups. Neuromuscular LBP characteristics were described with coordination, co-activation, trends, and fatigue measures. Altogether, a set of 327 secondary features was created where inputs into the classification models were further refined by employing neighborhood component analysis (NCA). NCA effectively reduced the number of features (<20 components), alongside preserving them in the original interpretable domain. A set of 23 different classifiers was employed and explored, resulting in classification accuracy of 0.94 for HS vs. LBP, 0.89 for HS vs. CLBP, 0.98 for HS vs. RLBP, and 0.89 for CLBP vs. RLBP differentiation. High median precision (0.97) and sensitivity (0.99) across all classifiers for HS vs. RLBP differentiation was obtained, with only three feature components utilized (out of 327). Support vector machines (SVM) and k-nearest neighbor (kNN) based classifiers consistently demonstrated best classification results. Different profiles of CLBP patients were presented and discussed. The suggested method demonstrated the potential for patients' subgrouping and subsequent more individualized rehabilitation treatments, backed by medical interpretations through feature modeling.

The Flexion Relaxation Phenomenon in Patients with Radiculopathy and Low Back Pain: A Cross-Sectional Study

“…Statistical analysis did not show any significant differences between both sides when analyzing myoelectric characteristics based on RMS calculation. It is worth noticing that such left–right differences can generally be captured by means of sEMG, especially for patients with radiculopathy, as presented in our work [ 33 ]. However, this comes at the expense of different sEMG measures being employed (e.g., signal slope change—SSC, permutation entropy—PE, or relative variance difference—RVD) and with more complex feature engineering and analysis involved.…”