Informatics in Medicine Unlocked (Jan 2021)
The role of frailty index in predicting readmission risk following total joint replacement using light gradient boosting machines
Abstract
Background: Total joint replacement (TJR) is a costly and prevalent surgery among older persons. Readmission is the most commonly experienced postoperative adverse event and is an important component to consider before surgery. A preoperative evaluation of postsurgical complications can assist with appropriate postoperative discharge decisions and can improve informed consent. Purpose: Our goal was to compare and use American Society of Anesthesiologists’ Physical Status Classification (ASA) versus modified frailty index (mFI-5) in readmission risk following TJR. Material & methods: We implemented light gradient boosting machine (LightGBM) as the main machine learning technique on data from American College of Surgeons National Surgical Quality Improvement Program (NSQIP) between 2011 and 2017 to preoperatively predict risk of postoperative readmission. We also compared results to extreme gradient boosting (XGB), logistic regression and random forests machine learning techniques. We included 22 presurgical diagnoses and symptoms as well as ASA and mFI-5 scores. However, we randomly split our cohort into 80% training and 20% hold out test data, performing 5-fold cross validation on the 80% training dataset. Model comparisons were based on area under the receiver operating characteristics curve (AUC) obtained on the hold out data. A variable importance analysis was performed on the best modelling technique. Results: Our analytic sample included 477, 616 surgeries. The readmission rate was 3.6% (n = 17 430). ASA class as a standalone predictor yielded an AUC of 0.607 and mFI-5 yielded an AUC of 0.585. LightGBM models using all the defined predictors resulted in an AUC of 0.67. Variable importance analysis on the final LightGBM model showed that ASA, age and mFI-5 are amongst the most important predictors. However, the frailty index can provide more detailed information on the requirement of care post-surgery. Conclusions: There is value in measuring not only the diagnoses and symptoms, but also patient frailty. As standalone variables, both ASA and mFI-5 are moderately associated with post TJR readmission. The LightGBM model predicted post-TJR readmission with moderate accuracy. Future studies should also include socioeconomic and behavioral factors as potential predictors of readmission.
