Explainable Machine Learning Applied to Bioelectrical Impedance for Low Back Pain: Classification and Pain-Score Prediction

TL;DR

This study uses machine learning with bioelectrical impedance to classify low back pain and predict pain scores, offering an objective and interpretable method.

Contribution

The novel application of explainable machine learning to bioelectrical impedance for low back pain classification and pain-score prediction.

Findings

The classifier achieved high discrimination between low back pain and healthy groups (ROC–AUC = 0.996).

Pain intensity prediction using VAS showed strong performance (R2 = 0.70).

SHAP values provided interpretable insights into feature importance and model behavior.

Abstract

(1) Background: Low back pain (LBP) is the most prevalent cause of disability worldwide, yet current assessment relies mainly on subjective questionnaires, underscoring the need for objective and interpretable biomarkers. Bioelectrical impedance parameter (BIP), quantified by resistance (R), impedance magnitude (Z), and phase angle (PA), reflects tissue hydration and cellular integrity and may provide physiological correlates of pain; (2) Methods: This cross-sectional study used lumbar BIP and demographic characteristics from 83 participants (38 with lumbar BIP and 45 normal controls). We applied Extreme Gradient Boosting (XGBoost), a regularized tree-based machine learning (ML) algorithm, with stratified five-fold cross-validation. Model interpretability was ensured using SHapley Additive exPlanations (SHAP), which provide global importance rankings and local feature attributions. Outcomes included classification of LBP versus healthy status and regression-based prediction of pain scales: the Visual Analog Scale (VAS), Oswestry Disability Index (ODI), and Roland–Morris Disability Questionnaire (RMDQ); (3) Results: The classifier achieved high discrimination (ROC–AUC = 0.996 ± 0.009, sensitivity = 0.950 ± 0.068, specificity = 0.977 ± 0.049). Pain prediction showed best performance for VAS (R2 = 0.70 ± 0.14; mean absolute error = 1.23 ± 0.27), with weaker performance for ODI and RMDQ; (4) Conclusions: These findings suggest that explainable ML models applied to BIP could discriminate between LBP and healthy groups and could estimate pain intensity, providing an objective complement to subjective assessments.

Figures