Development and Validation of a Predictive Model for Failure of Medical Management in Spinal Epidural Abscesses

BACKGROUND: The optimal management of spinal epidural abscesses (SEA) secondary to primary spinal infections has demonstrated large variability in the literature. Although some literature suggests a high rate of neurological deterioration, others suggest failure of medical management is uncommon. OBJECTIVE: To develop a predictive model to evaluate the likelihood of failure of medical therapy in the setting of SEA. METHODS: A retrospective review was conducted of all patients presenting with SEA from primary spinal infections. Patients presenting with MRI evidence of SEA without neurological deficits were included. Failure of medical management was defined as requiring surgical intervention over 72 hours after the initiation of antibiotics. A machine learning method (Risk-Calibrated Supersparse Linear Integer Model) was used to create a risk stratification score to identify patients at high risk for requiring surgical intervention. RESULTS: In total, 159 patients were identified as presenting with MRI findings of SEA without evidence of neurological deficit. Of these patients, 50 required delayed surgery compared with 109 whose infection were successfully treated with surgical intervention. The Spinal Epidural Abscess Predictive Score was created using a machine learning model with an area under the curve of 0.8043 with calibration error of 14.7%. Factors included active malignancy, spondylodiscitis, organism identification, blood cultures, and sex. The probability of failure of medical management ranged from 95% for a score of 7 or more. CONCLUSION: The Spinal Epidural Abscess Predictive Score model is a quick and accurate tool to assist in clinical decision-making in SEA.

Automated summary

A predictive model called the Spinal Epidural Abscess Predictive Score was developed using machine learning to assist in clinical decision-making for patients with spinal epidural abscesses. This study identified factors such as active malignancy, spondylodiscitis, organism identification, blood cultures, and sex that contribute to the likelihood of failure of medical management. The model had a high accuracy in predicting the need for surgical intervention, with a probability of failure of medical management reaching 95% for a score of 7 or more.