OBJECTIVES

The Centers for Medicare and Medicaid Services (CMS) is focusing on readmission within 30-days after total joint replacements for cost and quality considerations. This study evaluated machine learning (ML) algorithms for predicting the risk of 30-day readmission following hip and knee arthroplasty procedures.

METHODS

This study used data from the 2016 Nationwide Readmissions Database (NRD). Patients aged ³18 years with index admissions for elective primary Total Hip Arthroplasty (THA) or Total Knee Arthroplasty (TKA) were identified using the ICD-10-PCS codes,as primary procedure code. Models were trained on randomly partitioned 50% of the data, and the remaining dataset was used for validation and testing to predict readmission within 30 days of hospital discharge. The down-sampling technique was used on the training data for class imbalance and model hyperparameters were tuned using the 10-fold cross-validation. This study applied three ML algorithms to predict readmission, namely Least Absolute Shrinkage and Selection Operator (LASSO), Random forest (RF), and RuleFit. Model performance was evaluated using Area Under the Curve (AUC), accuracy, sensitivity, specificity, F1 score, and precision.

RESULTS

According to the NRD, there was a total of 417,211 patients with hospital admission for elective THA and TKA. Among these admissions, 14,208 (3.41%) patients were readmitted within 30 days of hospital discharge. The model performance from all algorithms was similar, with an AUC score of 0.61. The LASSO model had an accuracy of 0.66, while the RF model and RuleFit model had an accuracy of 0.65. The sensitivity, specificity, F1 score, and precision were also similar for three ML algorithms.

CONCLUSIONS

The three ML algorithms performed similarly based on the various model performances. The model performance measures should be carefully evaluated before applying these algorithms for evaluating the quality of care in hospitals.