Developing a Multivariable Prediction Model for Mortality in Intramedullary Spinal Cord Tumor Resection Patients

Introduction: Predictive models are instrumental in assessing mortality risk across various medical conditions, but few have been developed and validated for neurosurgical patients, especially those with rare pathologies. Intramedullary spinal cord tumors (IMSCTs), representing 2-4% of primary CNS neoplasms, have been inadequately studied due to their rarity. Operative decision-making and goals-of-care discussions with IMSCT patients are limited by lack of evidence predicting mortality. After assembling a large cohort of IMSCT patients spanning 20 years and multiple tumor pathologies, we developed a mortality prediction model for patients undergoing IMSCT resections.

Methods: Perioperative data were retrospectively collected on 138 patients with IMSCT resections performed at UPMC hospitals between December 2001 and January 2020. Logistic regression models were generated for mortality within the study period, 5 years of surgery, and 1 year of surgery. Predictors included tumor pathology, extent of resection, age at surgery, tumor location, and biological sex. Models were optimized via stepwise backward regression and evaluated for fit, discrimination, and parsimony via goodness-of-fit tests, likelihood ratio tests, and area under the curve (AUC) for receiver operating curves.

Results: Mortality within the study period was significantly associated with age at surgery (p=1.13 x10-6) and myxopapillary ependymoma pathology (p=0.04). The final model had an AUC value of 0.8360, indicating excellent discrimination between patients who did and did not die. Mortality within 5 years of surgery was significantly associated with only thoracolumbar location (p=0.02). The final model had an AUC value of 0.7141, indicating adequate discrimination. Finally, mortality within 1 year of surgery was significantly associated with myxopapillary ependymoma pathology (p=0.0087) and ependymoma pathology (p=0.01). The final model had an AUC value of 0.7928, indicating good discrimination.

Conclusion : These preliminary prediction models for mortality in IMSCT resection patients demonstrate how differently perioperative factors must be weighted in predicting mortality over distinct time periods. Although further refinement is required via meta-analyses and factor interaction testing, they represent an important step toward establishing accurate pre-operative mortality estimates for providers educating patients on IMSCT resection risks.