Development and validation of a predictive model for outcomes in shoulder arthroplasty: a multicenter analysis of nearly 2000 patients.

Overview

abstract

BACKGROUND: Guiding expectations following shoulder arthroplasty is important in improving patient satisfaction. The purpose of this study was to develop a predictive model to calculate 2-year American Shoulder and Elbow Surgeons (ASES) scores in shoulder arthroplasty patients from a comprehensive set of preoperative patient factors and types of arthroplasty performed. METHODS: This retrospective multicenter study included 1947 shoulder arthroplasties performed from 2010 to 2015 at 3 high-volume centers. Twenty-six variables were evaluated for an association with 2-year ASES scores, and variables with P < .20 in our pair-wise analysis were used to develop a predictive model. The prediction root-mean-square error was calculated. External validation was performed using data from 233 patients who underwent shoulder arthroplasty performed by a separate shoulder surgeon at a center not involved with creation of the predictive model. RESULTS: A total of 1947 patients were analyzed, and their data were used to construct the predictive model. Variables most associated with 2-year ASES scores were patient age, preoperative ASES score, disability, chronic obstructive pulmonary disease, alcohol use, anatomic vs. reverse total shoulder arthroplasty, and primary vs. revision shoulder arthroplasty. By use of cross validation, the prediction error was 20.1, the proportion of variance explained was 25.3%, the mean absolute error was 15.9, and the C statistic for the linear regression model was 0.66. After external validation, the mean difference between predicted and actual 2-year ASES scores was 12.7 points, within the accepted minimal clinically important difference after shoulder arthroplasty. DISCUSSION: Data from nearly 2000 shoulder arthroplasties allowed the development and validation of a model to predict 2-year ASES scores following shoulder arthroplasty. The model was accurate within the minimal clinically important difference in 85% of patients.