Machine learning-based identification and ranking of risk factors for lumbar paraspinal muscle atrophy.

BACKGROUND: Recent studies highlight the crucial role of the paraspinal musculature (PM), particularly the multifidus (MF), in spinal health and patient outcomes. However, factors associated with PM atrophy and their relative importance remain unclear. To address this gap, we analyzed factors linked to PM atrophy in patients undergoing lumbar fusion using machine learning, aiming to clarify the multifactorial mechanisms underlying this condition. METHODS: Fatty infiltration (FI) of the lumbar MF was measured as a proxy for muscular atrophy in patients undergoing lumbar spinal fusion. Two machine learning models, logistic regression and extreme gradient boosting (XGBoost), were trained to predict severe FI (> 50%) of the MF. Model performance was evaluated on unseen test data using receiver operating characteristic (ROC) analysis and Brier score, and predictor importance was assessed via SHAP (SHapley Additive exPlanations). RESULTS: The study included 316 patients, primarily treated due to lumbar spinal stenosis. Both machine learning models effectively predicted severe MF atrophy, with an area under the curve (AUC) of 0.83 (95% CI 0.74–0.83) for the logistic regression and 0.88 (95% CI 0.81–0.88) for the XGBoost. In the logistic regression model, only sex, age, and facet joint degeneration were significant predictors. The XGBoost model identified the same top three variables, while the lumbar endplate score and bone mineral density ranked higher than in logistic regression. CONCLUSION: This study introduces a novel framework for analyzing factors influencing PM atrophy, highlighting the intricate interplay between demographic variables like age and sex and facet joint degeneration. By applying modern machine learning techniques, we improved predictive accuracy and identified endplate and bone changes as strongly associated factors, offering valuable insights into the mechanisms shaping muscle health in lumbar conditions.