P308. Utilizing machine learning for enhanced surgical outcomes: a national analysis of intramedullary spinal cord tumors

• Published in: The spine journal Vol. 24; no. 9; pp. S213 - S214
• Main Author: Tsai, Sung Huang Laurent
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.09.2024
• ISSN: 1529-9430
• DOI: 10.1016/j.spinee.2024.06.431

Question: This study sought to identify impactful variables on health care outcomes in patients undergoing surgery for intradural intramedullary spinal cord tumors (IMSCTs). Objectives included developing risk calculators for mortality and extended length of stay (eLOS), stratifying IMSCTs by histology, and evaluating potential interventions to mitigate adverse outcomes. Findings: Analysis of 7,243 IMSCT cases over 12 years revealed evolving management strategies for astrocytomas, ependymomas, and hemangioblastomas. Machine learning models, yielding AUCs of 0.721 for mortality and 0.586 for eLOS, identified critical predictors including patient age, year of diagnosis, clinical behavior, histology, radiation therapy, insurance status, distance to hospital, tumor grade, length of stay, extent of tumor resection, tumor size, and sex. Additionally, web-based tools for mortality and eLOS prediction were successfully developed and deployed. Meaning: This comprehensive national study offers valuable insights into the outcomes of IMSCTs, highlighting significant shifts in tumor management and the effectiveness of machine learning in predictive modeling. The identified predictive factors are instrumental in facilitating early intervention and targeted care strategies, potentially enhancing patient outcomes and optimizing health care resource utilization. This study investigates health care outcomes in patients undergoing surgical resection for intradural intramedullary spinal cord tumors (IMSCTs), employing the National Cancer Data Base (NCDB) to identify key variables. We aimed to develop supervised machine learning-based risk calculators to predict high-risk patients for mortality and extended length of stay (eLOS), stratifying IMSCTs by histology to enhance understanding and guide intervention strategies for adverse outcomes. A retrospective, multicenter cohort analysis of patients with surgically-treated IMSCTs (2004-2017) was conducted using the NCDB. We extracted demographic and comorbidity data, employing descriptive statistics and supervised machine learning algorithms to predict mortality and eLOS. The National Cancer Database (NCDB) is one of the largest cancer registries in the United States and contains almost 34 million cases from over 1500 hospitals. Data are collected from selected health registries accredited by the American College of Surgeons’ Commission on Cancer (https://www.facs.org/quality%20programs/cancer/coc). A retrospective cohort analysis using the NCDB was queried for patients diagnosed with and surgically treated for spinal cord intramedullary spinal cord tumors (IMSCTs): astrocytomas, ependymomas, and hemangioblastomas from January 1, 2004, through December 31, 2017. We used the ICD-O-3 histological codes designating astrocytomas (9382, 9384, 9400, 9401, 9410, 9411, 9412, 9420, 9421, 9505), ependymomas (9383, 9391, 9392, 9393, 9394) and hemangioblastomas (9161) and ICD-O-3 topographical codes for the spinal cord (C72.0), and cauda equina (C72.1) to identify pertinent cases. This study follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.It was determined to be exempt by the Mayo Clinic institutional review board (IRB) and did not require informed consent to use de-identified data. Surgical cases at a distant site were excluded to avoid confounding different surgical procedures and ensure that the surgical procedures were on the primary site The primary outcomes of interest were 5-year overall mortality and extended length of stay (eLOS), defined as greater than or equal to the 75th percentile (6 days). Baseline patient characteristics, hospital characteristics, and tumor characteristics were analyzed for each group. Data Source and Patient Cohort The National Cancer Database (NCDB) is one of the largest cancer registries in the United States and contains almost 34 million cases from over 1,500 hospitals. Data are collected from selected health registries accredited by the American College of Surgeons’ Commission on Cancer (https://www.facs.org/quality%20programs/cancer/coc). A retrospective cohort analysis using the NCDB was queried for patients diagnosed with and surgically treated for spinal cord intramedullary spinal cord tumors (IMSCTs): astrocytomas, ependymomas, and hemangioblastomas from January 1, 2004, through December 31, 2017. We used the ICD-O-3 histological codes designating astrocytomas (9382, 9384, 9400, 9401, 9410, 9411, 9412, 9420, 9421, 9505), ependymomas (9383, 9391, 9392, 9393, 9394) and hemangioblastomas (9161) and ICD-O-3 topographical codes for the spinal cord (C72.0), and cauda equina (C72.1) to identify pertinent cases. This study follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.13 It was determined to be exempt by the Mayo Clinic institutional review board (IRB) and did not require informed consent to use de-identified data. Surgical cases at a distant site were excluded to avoid confounding different surgical procedures and ensure that the surgical procedures were on the primary site. The demographic and clinical characteristics of included individuals are presented in Primary and Secondary Outcome Variables: The primary outcomes of interest were 5-year overall mortality and extended length of stay (eLOS), defined as greater than or equal to the 75th percentile (6 days). Baseline patient characteristics, hospital characteristics, and tumor characteristics were analyzed for each group. Statistical Analysis: Descriptive statistics characterized the cohort. Continuous variables were compared using Student's t-test and Mann–Whitney tests, while categorical variables utilized Chi-squared and Fisher's exact tests. Trend slopes in IMSCT treatments and length of stay were analyzed using piecewise joinpoint regression and the Mann-Kendall test. Data partitioning (70%/15%/15% for train/validation/test sets) and class imbalance addressed through SMOTE. Hyperparameter tuning and model evaluation encompassed an array of supervised machine learning algorithms, with final models refined by explainable AI techniques, including SHAP analysis. The resulting algorithms for mortality and eLOS prediction were integrated into a publicly accessible web-based application. Statistical significance was set at p < 0.05, with analyses conducted using R and Python. The study encompassed 7,243 surgically treated IMSCT cases, including 612 astrocytomas (8.5%), 6,041 ependymomas (83.4%), and 590 hemangioblastomas (8.1%). Mortality and eLOS rates were observed at 10.2% and 27.1%, respectively. Over 12 years (2004-2016), significant management shifts were noted for these spinal tumor types. The predictive models achieved AUCs of 0.721 for mortality and 0.586 for eLOS. Key predictive features for mortality included age, diagnosis year, behavior, histology, radiation, insurance status, patient-hospital distance, tumor grade and size, length of stay, subtotal resection (STR) to gross total resection (GTR), and sex. For eLOS, additional predictors were diagnosis-surgery interval, Charlson/Deyo score, and surgical approach. Web-based tools for both outcomes have been deployed: https://imsct-elos-predict.herokuapp.com/; https://imsct-risk-calcualor.herokuapp.com/ Our nationwide analysis underscores the evolution in IMSCT management and demonstrates the efficacy of machine learning in predicting mortality and eLOS, providing valuable insights for improved patient care. This abstract does not discuss or include any applicable devices or drugs.