Secondary spinal epidural abscess represents a severe and potentially catastrophic complication in patients diagnosed with pyogenic spondylitis. The expanding infection can rapidly compress the spinal cord or the cauda equina, precipitating irreversible neurological deficits such as paraplegia. According to the research team, once an epidural abscess has formed, patients often require emergency surgical decompression to relieve spinal cord compression, and the condition remains closely tied to significantly poorer overall outcomes. Consequently, developing reliable tools for the early identification of individuals at high risk for abscess progression before physical abscess formation occurs carries substantial clinical value for inpatient monitoring and timely intervention.
Within this total cohort of 240 patients, 60 individuals, representing exactly 25.0 percent of the study population, subsequently developed a secondary spinal epidural abscess.
The investigative team initially assembled a comprehensive set of 57 candidate variables derived from clinical findings, laboratory assays, and imaging examinations that were first acquired early after patient admission. In the training cohort, candidate predictors were systematically screened using a two-stage feature selection pipeline. The researchers first conducted univariate logistic regression, screening out variables that failed to meet a significance threshold of P < 0.10. Surviving variables were then subjected to least absolute shrinkage and selection operator regression, commonly known as LASSO regression, utilizing 10-fold cross-validation to penalize redundant parameters.
Through this statistical screening, six final clinical predictors were identified: intervertebral disc T2 hyperintensity or enhancement, abnormal sensory level, number of vertebrae with endplate destruction, maximum diameter of paravertebral abscess, procalcitonin, and Frankel grade E (a protective factor).
Using these six predictors, the investigators trained and compared four machine learning algorithms: L2-regularized logistic regression, random forest, support vector machine, and eXtreme Gradient Boosting, also designated as XGBoost. When evaluated on the 72 patients in the held-out test cohort, the random forest model demonstrated the numerically highest discriminative capacity, achieving an area under the receiver operating characteristic curve of 0.871, with a 95 percent confidence interval ranging from 0.760 to 0.982. The random forest model also attained an area under the precision-recall curve of 0.728. When operating at an optimized risk probability cutoff threshold of 0.312, the model demonstrated an overall classification accuracy of 79.2 percent, correctly identifying 83.3 percent of patients who developed an abscess (sensitivity) and 77.8 percent of patients who did not (specificity).
The authors also evaluated the statistical calibration and internal reliability of the random forest predictions. Calibration analysis revealed close concordance between estimated probabilities and observed disease outcomes, reflected by a low Brier score of 0.108 and a non-significant Hosmer-Lemeshow goodness-of-fit test statistic with a chi-square of 6.890 and a P value of 0.549. Internal validation conducted via 1,000 bootstrap resamples confirmed the stability of the model, yielding an optimism-corrected area under the curve of 0.872. Furthermore, in three distinct sensitivity analyses designed to assess performance across varying analytical conditions, the model’s discriminative ability remained stable, maintaining area under the curve values bounded between 0.851 and 0.868.
To assess clinical utility and interpretability, the study incorporated decision curve analysis alongside SHapley Additive exPlanations, known as SHAP analysis. The decision curve evaluation demonstrated positive clinical net benefit across an operational threshold probability range of 0.05 to 0.70, indicating that applying the model offers practical clinical value across diverse decision-making thresholds. The investigators noted that this random forest framework provides an objective, quantitative tool for early risk stratification and surveillance planning among hospitalized pyogenic spondylitis patients. However, the authors explicitly emphasized that because the investigation was conducted retrospectively at a single hospital center, prospective multicenter studies will be necessary to further validate the model’s generalizability and real-world performance before widespread clinical adoption.
Subject of Research: Machine learning risk prediction for secondary spinal epidural abscess in pyogenic spondylitis
Article Title: Development and validation of an early warning model for secondary spinal epidural abscess in pyogenic spondylitis based on interpretable machine learning: a single-center retrospective cohort study
Article References: Liu, J., Zhang, X., Tan, J., Tang, J., & Jiang, B. (2026). Development and validation of an early warning model for secondary spinal epidural abscess in pyogenic spondylitis based on interpretable machine learning: a single-center retrospective cohort study. Frontiers in Public Health, 14, Article 1979233. https://doi.org/10.3389/fpubh.2026.1979233
Image Credits: AI Generated
DOI: 10.3389/fpubh.2026.1979233
Keywords: pyogenic spondylitis, spinal epidural abscess, machine learning, random forest, early warning model, procalcitonin, intervertebral disc T2 hyperintensity, abnormal sensory level, endplate destruction, paravertebral abscess, Frankel grade E, risk stratification, retrospective cohort study, decision curve analysis, SHAP interpretability, LASSO regression, SMOTE, internal validation
News Source: Ophelia Keating. (November 14, 2026). Random Forest Model Predicts Secondary Spinal Epidural Abscess in Pyogenic Spondylitis. Scienmag.



