P83. Predictive modeling for pseudarthrosis performance benchmarking in 404 patients with a minimum two-year follow up

Abstract

BACKGROUND CONTEXT In the past, averages from published studies have been used for performance benchmarking. However, in practice and in large studies, usually only a small subset of patients fall within the “average” and therefore the average may not provide a valid benchmark rate. One can overperform or underperform the average and still attain the predicted performance benchmark. Predictive modeling can be used to help set performance standards based on a large number of variables including patients' specific factors, procedure details, use of osteogenic products, site specific variations and many others. This study attempts to use predictive modeling for pseudarthrosis performance benchmarking as a proof of concept in ASD surgery. PURPOSE To use predictive modeling for performance benchmarking in high-volume centers and demonstrate that it is more appropriate than simply using averages. STUDY DESIGN/SETTING Retrospective review of a prospective, multicenter ASD database from 11 different sites PATIENT SAMPLE ASD operative patients with age ≥18, Coronal Cobb ≥ 20deg, SVA ≥5cm, PT≥25deg, and/or thoracic kyphosis (TK) ≥ 60deg with min 2-year follow-up. Exclusion criteria: having a revision for any indication other than pseudo in order to reduce confounding of potential pseudo as a result of the revision surgery. OUTCOME MEASURES HRQOL scores: Oswestry Disability Index (ODI), Short form-36 (SF36), Scoliosis Research Society (SRS22), back/leg pain numerical rating scale (NRS). Radiographic values: max coronal cobb angle, coronal C7 plumb line, pelvic tilt (PT), mismatch between pelvic incidence and lumbar lordosis (PI-LL), thoracic kyphosis (TK), C7 sagittal vertical axis (SVA). Posterior fusion was graded to determine if pseudo has occurred within 2 years postoperative. Demographic, frailty, surgical including BMP and complications data were also collected. METHODS A prior validated and published pseudo predictive model has been constructed from 336 ASD patients with an accuracy of 91.3% and an AUC of 0.94 using 21 out of 82 total variables listed above. This model was deployed with an updated set of patients to determine the predicted pseudo rate for each individual surgical site. Actual pseudo rates of the 11 contributing surgical sites were compared to the predicted values as a means to assess for performance benchmarking. RESULTS A total of 403 patients were included (80.1% Female, avg age 57.9±14.9 years) from a total of 502 operative patients with 99 excluded for having a revision for indication other than pseudo. A total of 129 (32.0%) had pseudo by 2 years. The overall pseudo rates per year were the following: 2008-20.0%, 2009-37.7%, 2010-31.4%, 2011-29.3%, 2012-36.7%, 2013-32.0%, 2014-26.6%, 2015-38.1%. Six sites had actual rates above the overall rate of 32% with 5 sites below. However, the predicted rates varied according to each site and included rates above/below the overall rate. All of the actual rates were larger than the predicted rates except for 4 sites in which the actual and predicted rates were the same. CONCLUSIONS The pseudo rate varied per year and per site. Predictive modeling was able to provide a customized pseudoarthosis rate for each site considering multiple variables allowing for performance benchmarking instead of the average. Even though a site was above the overall average rate, they may be predicted to have a higher rate given the type of patients being treated or procedures done at that site. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.