2021 ISAKOS Biennial Congress Paper
Predicting Factors For Achieving Minimally Clinical Important Difference After Primary Shoulder Arthroplasty: A Machine Learning Model
Ayoosh Pareek, MD, Rochester, MN UNITED STATES
Micah Nieboer, MD, Rochester, MN UNITED STATES
Jianing Man, PhD, Rochester, MN UNITED STATES
Ronda Esper, BS, Rochester, MN UNITED STATES
Kalyan Pasupathy, PhD, Rochester, MN UNITED STATES
Joaquin Sanchez-Sotelo, MD, Rochester, MN UNITED STATES
Mayo Clinic, Rochester, MN, UNITED STATES
FDA Status Not Applicable
Machine learning methodology identified age, BMI, and forward flexion as the most important factors in prediction of MCID after shoulder arthroplasty
Introduction/Purpose: Previous studies have tried to predict minimally clinical important difference (MCID) after total or reverse shoulder arthroplasty (TSA, RSA). However, they have been limited by either small sample sizes or lack of detail on the accuracy of their predictive results. The purpose of this study was to use machine learning to develop a predictive model for achieving MCID after TSA and RSA considering demographic, psychosocial, and physical exam factors.
All patients who underwent primary TSA or RSA by a single surgeon with preoperative and 1-year postoperative ASES scores were evaluated to determine whether they had achieved maximal clinical benefit from the procedures. Patients with complications or reoperations from the surgery within the first year were excluded due to obvious effect on ASES scores. The study population included 166 patients (49% male) that had undergone TSA (36%) or RSA (64%) with a mean age of 70.4 (SD 8.8).Data collected included patient demographics (age, BMI, gender, diabetes and other), psychosocial factors (tobacco use, mental health disorders), physical exam parameters type of implant, and indication for arthroplasty. Data was randomly divided into two sets (80% for training and 20% for testing) and various machine learning algorithms were compared (Neural Network, Regression Tree, XGBoost, and Random Forest). The XGBoost ensemble method had the highest accuracy and was chosen.
Overall, the mean preoperative to postoperative ASES score change was 40.5 points (SD 22.9, p<0.001) and 74% patients achieved MCID. A machine learning model using the above methodology was created using two discrete steps. The first model was built using all parameters except for preoperative ASES score. The second model was built using imaging related parameters in addition to preoperative ASES scores to evaluate the increased accuracy in prediction. For the first model, the four most important variables were age at surgery, BMI, preoperative external rotation, and preoperative forward flexion (Figure 1). The overall area under the curve (AUC) of the test data model was 80%, which was deemed to be a very good model (Figure 2). The second model including preoperative ASES scores had an increased accuracy of testing data of 85%. In addition, this model did not rely on the same variables, as ASES became the most important variable in addition to BMI, age, and preoperative forward flexion (Figure 3).
In our study, 75% of the patients who had undergone primary shoulder arthroplasty had achieved MCID at one year. Machine learning methodology identified age, BMI, and forward flexion as the most important factors in prediction of MCID. The addition of preoperative ASES scores appears to improve model predictability when predicting which patients will achieve MCID.