166 Predictive Modeling of Length of Hospital Stay Following Adult Spinal Deformity Correction: Analysis of 653 Patients With an Accuracy of 75% Within 2 Days. (August 2016)

• Record Type: Journal Article
• Authors: Scheer, Justin K.; Ailon, Tamir T.; Smith, Justin S.; Hart, Robert; Burton, Douglas C.; Bess, Shay; Neuman, Brian J.; Passias, Peter G.; Miller, Emily; Shaffrey, Christopher I.; Schwab, Frank; Lafage, Virginie; Klineberg, Eric; Ames, Christopher P.
• Is Part Of: Clinical neurosurgery. Volume 63(2016)Supplement 1
• Publication Date: 2016-08
• Subjects: Nervous system -- Surgery -- Congresses; Neurosurgery; Nervous system -- Surgery; Neurologie; Congresses; Conference papers and proceedings; 617.48
• DOI: 10.1227/01.neu.0000489735.46846.2b
• Languages: English
• ISSNs: 0069-4827

Abstract

Abstract : INTRODUCTION: The length of stay (LOS) following adult spinal deformity (ASD) surgery is a critical time period allowing for recovery to levels safe enough to return home or to rehabilitation. Thus, the goal is to minimize it for conserving hospital resources and third-party payer pressure. Factors related to LOS have not been studied nor has a predictive model been created. The goal of this study was to construct a preadmission predictive model based on patients' baseline variables and modifiable surgical parameters. METHODS: Retrospective review of a multicenter, prospective ASD database. Inclusion criteria: operative patients, age >18 years, ASD. Patients with staged surgery at a separate hospitalization or LOS >30 days were excluded. Sixty-six variables were initially evaluated with 40 being used for model building following univariable predictor importance = 0.90, redundancy, and collinearity testing. Variables included: demographics, comorbidities, preoperative health-related quality of life, preoperative coronal and sagittal radiographic parameters, and modifiable surgical factors (Figure). A generalized linear model was constructed by using a training data set developed from a bootstrapped sample with replacement using a random number generator. Patients randomly omitted from the bootstrapped sample composed the testing data set. Accuracy was calculated by comparison of predicted LOS with the actual LOS. RESULTS: A total of 689 patients were eligible; 653 met inclusion criteria. The mean LOS was 7.9 ± 4.1 days (range: 1-28). Following bootstrapping, 893 patients were modeled in total, Training: 653, Testing: 240 (36.6%). The linear correlations for the training and testing data sets were 0.632 and 0.507, respectively. Testing dataset accuracy within 2 days of actual LOS was 75.4% (181/240 patients). CONCLUSION: A successful model was created to predict LOS to an accuracy of 75% within 2 days. There are some factors related to LOS that are not likely captured in large databases, which may partially explain the 75% accuracy, such as rehabilitation bed availability and social support resources.