Objective <p>Based on multicenter clinical data, this study aimed to develop and validate a predictive model for chronic low back pain (CLBP) after lumbar decompression surgery in patients with diabetes mellitus and lumbar disc herniation. The model integrated metabolic indicators and imaging-derived features of the paraspinal muscles.</p> Methods <p>This study was designed as a multicenter retrospective cohort study. A total of 2776 patients with diabetes mellitus and lumbar disc herniation were included. All patients underwent unilateral biportal endoscopic (UBE) decompression between January 2021 and December 2024 across six medical centers. Postoperative CLBP was defined as persistent or recurrent low back or lumbosacral pain lasting more than six months after surgery. A visual analogue scale (VAS) score greater than 4 was used as the threshold for outcome assessment. Patients were excluded if postoperative pain could be attributed to structural complications or recurrent pathology. Patients with severe psychiatric or psychosocial conditions that might impair the reliability of pain assessment were also excluded. Data from four centers were used for model development and internal validation, while the remaining two centers served as independent external validation cohorts. Clinical variables, laboratory findings, and imaging parameters were integrated to develop multiple machine learning-based predictive models. Model performance was evaluated based on discrimination, calibration, and clinical net benefit. In addition, restricted cubic spline (RCS) analysis was performed to assess potential nonlinear relationships between key continuous variables and postoperative CLBP risk. Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) were used to interpret model predictions at both global and individual levels.</p> Results <p>Among individual models, the ExtraTrees algorithm achieved the highest discriminative performance in the validation cohort (AUC = 0.855) and showed stable performance across the two external validation cohorts. However, its sensitivity (0.413) and F1 score (0.531) remained relatively low. After comprehensive evaluation of discrimination, classification performance, and calibration, a stacking ensemble model combining XGBoost, artificial neural networks (ANN), and linear discriminant analysis (LDA) was selected as the final model due to its more balanced overall performance. In the validation cohort, this model achieved an AUC of 0.838, which was slightly lower than that of ExtraTrees. However, it demonstrated substantial improvements in sensitivity, F1 score, and accuracy by 43.8%, 24.7%, and 4.9%, respectively, along with improved calibration performance (Brier score = 0.146). Decision curve analysis demonstrated stable clinical net benefit across the validation cohort and both external validation cohorts. RCS analysis revealed significant nonlinear associations between postoperative CLBP risk and age, admission blood glucose, glycated hemoglobin, psoas muscle index, multifidus fat infiltration, albumin, alkaline phosphatase, and serum calcium. SHAP and LIME analyses identified intervertebral disc degeneration grade, age, psoas muscle index, admission blood glucose, and paraspinal muscle fat infiltration as the most important predictors in the model.</p> Conclusion <p>This multicenter study developed and externally validated a predictive model for postoperative CLBP using integrated clinical and imaging data. The model demonstrated good discriminative ability and interpretability. It may be useful for estimating the risk of CLBP after lumbar decompression surgery in patients with diabetes mellitus. Metabolic abnormalities and paraspinal muscle degeneration-related features contributed substantially to model predictions, suggesting a statistical association with postoperative chronic pain development. The model may support individualized risk stratification, postoperative follow-up, and rehabilitation planning. However, given its limited sensitivity, further validation in prospective studies is required before clinical application.</p> <p><b>Keywords:</b>Lumbar Disc Herniation; Diabetes Mellitus; Unilateral Biportal Endoscopy; Prediction Model; Machine Learning; StackingModel</p>