Machine Learning-Based Predictive Models for 90-Day Readmission of Total Joint Arthroplasty Using Comprehensive Electronic Health Records and Patient-Reported Outcome Measures

Abstract

Background: The Centers for Medicare & Medicaid Services currently incentivizes hospitals to reduce postdischarge adverse events such as unplanned hospital readmissions for patients who underwent total joint arthroplasty (TJA). This study aimed to predict 90-day TJA readmissions from our comprehensive electronic health record data and routinely collected patient-reported outcome measures.

Methods: We retrospectively queried all TJA-related readmissions in our tertiary care center between 2016 and 2019. A total of 104-episode care characteristics and preoperative patient-reported outcome measures were used to develop several machine learning models for prediction performance evaluation and comparison. For interpretability, a logistic regression model was built to investigate the statistical significance, magnitudes, and directions of associations between risk factors and readmission.

Results: Given the significant imbalanced outcome (5.8% of patients were readmitted), our models robustly predicted the outcome, yielding areas under the receiver operating characteristic curves over 0.8, recalls over 0.5, and precisions over 0.5. In addition, the logistic regression model identified risk factors predicting readmission: diabetes, preadmission medication prescriptions (ie, nonsteroidal anti-inflammatory drug, corticosteroid, and narcotic), discharge to a skilled nursing facility, and postdischarge care behaviors within 90 days. Notably, low self-reported confidence to carry out social activities accurately predicted readmission.

Conclusions: A machine learning model can help identify patients who are at substantially increased risk of a readmission after TJA. This finding may allow for health-care providers to increase resources targeting these patients. In addition, a poor response to the "social activities" question may be a useful indicator that predicts a significant increased risk of readmission after TJA.

Keywords: Electronic health records; Hospital readmission; Machine learning; Patient-reported outcome measures; Total joint arthroplasty.

Figure 1

The list of the 10 variable categories and the timeline of the care categories. The categories related to the timeline were displayed in the first row, and the remaining categories were present in the second row. Medication prescription for pain management was defined as the medication being prescribed between 1 and 3 months prior to hospitalization, whereas one for surgery preparation was defined as the medication being prescribed within 1 month prior to hospitalization. CAM, confusion assessment method; DBP, diastolic blood pressure; ER, emergency room; HOOS, hip injury disability and osteoarthritis outcome score; KOOS, Knee Injury and Osteoarthritis Outcome score; MAP, mean arterial pressure; NSAID, nonsteroidal anti-inflammatory drugs; OP, operation; PROMIS, Patient-Reported Outcomes Measurement Information System; RAPT, Risk Assessment and Prediction Tool; SBP, systolic blood pressure.

Figure 2

Forest plot of the risk factors identification associated with 90-day readmission. An odds ratio (OR) and its 95% confidence interval (CI) for each risk factor are displayed in the middle. An arrow indicates a 95% CI that exceeds the limit of the chart (the left limit = 0.37 and the right limit = 12.18). The corresponding statistics are present in the right columns. A P-value less than .05 indicates a significant risk factor. PROMIS-10 Global09 states, “In general, please rate how well you carry out your usual social activities and roles.” ER, emergency room; NSAID, nonsteroidal anti-inflammatory drugs; PROMIS, patient-reported outcome measurement information system.