What predicts prolonged length of stay after surgery for cervical spondylotic myelopathy? A QOD CSM study

Cervical spondylotic myelopathy (CSM) is a leading cause of spinal cord dysfunction requiring surgical intervention. Prolonged length of stay (LOS) after CSM surgery is associated with worse outcomes, increased complications, greater financial burden, and inefficient resource utilization. This study aimed to develop machine learning models to predict prolonged LOS after CSM surgery and to improve patient counseling, perioperative optimization, and discharge planning.

The authors analyzed prospectively collected data from 14 high-accruing Spine CORe™ sites in the Quality Outcomes Database (QOD) of adult patients who underwent elective surgery for CSM. Patients with missing data were excluded. Machine learning models were trained to predict prolonged LOS, defined as ≥ 3 days. Models incorporated a wide range of preoperative demographic, clinical, and surgical variables. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC) analysis, and significant predictors were extracted from the logistic regression model.

Of the 1141 patients identified as having undergone elective surgery for CSM, 1020 were included. Logistic regression, XGBoost, and random forest models demonstrated excellent performance with mean ± standard deviation AUROC values of 0.88 ± 0.02, 0.89 ± 0.02, and 0.89 ± 0.02, respectively. Prior shoulder surgery (OR 2.33, 95% CI 2.16-2.51, p = 0.04), greater total fused levels (OR 1.89, 95% CI 1.83-1.95, p < 0.001), and diabetes (OR 1.75, 95% CI 1.66-1.83, p = 0.049) were predictors of prolonged LOS. In contrast, anterior surgical approach (OR 0.14, 95% CI 0.12-0.17, p < 0.001), radicular motor deficit (OR 0.49, 95% CI 0.45-0.54, p = 0.030), and greater baseline modified Japanese Orthopaedic Association (mJOA) score (OR 0.87, 95% CI 0.86-0.88, p = 0.007) were associated with a lower likelihood of prolonged LOS. Subgroup analyses revealed that patients with radicular motor deficit and prior shoulder surgery differed demographically, clinically, and surgically compared to those without.

In this large cohort of patients operated on for CSM, prior shoulder surgery, greater number of levels fused, and diabetes were significant positive predictors of prolonged LOS, while anterior approach surgery, radicular motor deficit, and higher baseline mJOA scores were predictive of a shorter inpatient stay. Patients with prior shoulder surgery and those with radicular motor deficit may represent distinct and clinically important subgroups within the broader CSM population. Machine learning models demonstrated excellent performance for predicting prolonged LOS from purely preoperative variables. The findings of this study may help enhance preoperative counseling, perioperative care, and resource utilization for CSM surgery.