Investigating Predictors of Increased Length of Stay After Resection of Vestibular Schwannoma Using Machine Learning. Issue 5 (June 2021)

• Record Type: Journal Article
• Authors: Dang, Sabina; Manzoor, Nauman F.; Chowdhury, Naweed; Tittman, Sarah M.; Yancey, Kristen L.; Monsour, Meredith A.; O'Malley, Matthew R.; Rivas, Alejandro; Haynes, David S.; Bennett, Marc L.
• Is Part Of: Otology & neurotology. Volume 42:Issue 5(2021)
• Publication Date: 2021-06
• Subjects: Acoustic neuroma -- Cost effectiveness -- Length of stay -- Machine learning -- Skull base; Otology -- Periodicals; Ear -- Diseases -- Periodicals; Skull base -- Surgery -- Periodicals; 617.8005
• DOI: 10.1097/MAO.0000000000003042
• Languages: English
• ISSNs: 1531-7129

Abstract

Abstract : Objective: To evaluate the predictors of prolonged length of stay (LOS) after vestibular schwannoma resection. Study Design: Retrospective chart review. Setting: Tertiary referral center. Patients: Patients who underwent vestibular schwannoma resection between 2008 and 2019. Interventions: Variables of interest included age, body mass index, comorbidities, symptoms, previous intervention, microsurgical approach, extent of resection, operative time, preoperative tumor volume, and postoperative complications. Predictive modeling was done through multivariable linear regression and random forest models with 80% of patients used for model training and the remaining 20% used for performance testing. Main Outcome Measures: LOS was evaluated as the number of days from surgery to discharge. Results: Four hundred one cases from 2008 to 2019 were included with a mean LOS of 3.0 (IQR = 3.0–4.0). Postoperatively, 14 (3.5%) of patients had LOS greater than two standard deviations from the mean (11 days). In a multivariate linear regression model (adjusted R 2 = 0.22; p < 0.001), preoperative tumor volume ( p < 0.001), coronary artery disease ( p = 0.002), hypertension ( p = 0.029), and any major complication ( p < 0.001) were associated with increased LOS (by 0.12, 3.79, 0.87, and 3.20 days respectively). A machine learning analysis using a random forest identified several potential nonlinear relationships between LOS and preoperative tumor dimensions (length, volume) and operative time that were not captured on regression. The random forest model had lower prediction error compared to the regression model (RMSE 5.67 vs. 44.59). Conclusions: Tumor volume, coronary artery disease, hypertension, and major complications impact LOS. Machine learning methods may identify nonlinear relationships worthy of targeted clinical investigation and allow for more accurate patient counseling. Abstract : Supplemental Digital Content is available in the text