Identifying Predictors of Adverse Outcomes and Developing Risk Calculators for Intramedullary Spinal Cord Tumor Patients Undergoing Surgery: A National Analysis Using Machine Learning.

Introduction: Intradural intramedullary spinal cord tumors (IMSCTs) significantly burden the healthcare system due to their complicated surgical management. Using the NCDB database, we aimed to identify variables that provide insights into healthcare outcomes for patients undergoing surgical resection of IMSCTs and to develop risk calculators using supervised machine learning to detect patients at high risk of extended length of stay (eLOS). The study also aims to stratify IMSCTs based on histology to enhance the understanding of factors contributing to adverse outcomes and to develop interventions to reduce their occurrence.

Methods: We conducted a retrospective, multicenter cohort analysis of patients diagnosed with and surgically treated for IMSCTs between January 1, 2004, and December 31, 2017, using the NCDB database. Demographics and comorbidities were extracted. Descriptive statistics were used to define the overall study cohort. Supervised machine learning algorithms were trained to predict the primary outcome: eLOS.

Results: Our study analyzed 7,243 cases of surgically treated intramedullary spinal cord tumors (IMSCT), 612 cases of astrocytomas (8.5%), 6,041 cases of ependymomas (83.4%), and 590 cases of hemangioblastomas (8.1%). The observed rate of mortality and eLOS was 10.2% and 27.1%, respectively. The model performed with an AUC of 0.721 and 0.586 for mortality and eLOS, respectively. The top features utilized by the models to predict mortality include age, year of diagnosis, behavior, histology, radiation, insurance status, the distance between the patient's residence and hospital, grade, length of stay, GTR, STR, tumor size, and sex. Regarding eLOS, the top predictors were similar to the above, with the addition of days between diagnosis and surgery, Charlson/Deyo score, and surgical approach. Web-based tools for both outcomes were deployed: https://imsct-elos-predict.herokuapp.com/

Conclusion : In this study, we provide a nationwide analysis of patients surgically treated for IMSCTs. Our study highlights the main shifts in managing IMSCTs over a 12-year period, focusing on ependymomas, astrocytomas, and hemangioblastomas. We also demonstrate the potential of machine learning models in predicting mortality and extended length of stay in IMSCT patients.