Development and Clinical Utility of Machine Learning Models for Prediction of Same‐Day Discharge in Outpatient Hip and Knee Replacement: A Prognostic Study
Christoffer C. Jørgensen, Jakob B. Frederiksen, Henrik Kehlet, Martin Lindberg‐Larsen, Kirill Gromov, Claus Varnum, Thomas Jakobsen, Manuel Josef Bieder, Mikkel Rathsach Andersen, Søren Overgaard, Torben B. Hansen, Troels PetersenABSTRACT
Background
Hip and knee replacement are common procedures with an increasing focus on same‐day surgery. However, capacity constraints limit the number of eligible patients actually being scheduled for same‐day discharge, calling for further selection of those with the highest likelihood of same‐day discharge.
Methods
A prognostic study from September 2022 to April 2024 aiming to develop and evaluate three machine learning models of increasing complexity for prediction of successful same‐day discharge after hip and knee replacement. Data was collected from six Danish departments with similar same‐day surgery protocols and same‐day surgery eligibility was according to predefined clinical criteria. The models were evaluated using receiver operating characteristic and clinical utility curves depicting the potential increase in same‐day discharge at different same‐day surgery capacities.
Results
Of 5387 eligible patients, 4466 (82.9%) were scheduled for same‐day surgery. Of these, 3085 (69.1%) achieved same‐day discharge and 1381 (30.9%) were admitted. The remaining 921 (17.1%) were planned as in‐patients. The area under the receiver operating curve showed poor but marginally increasing predictive ability (0.586, 0.602, and 0.603, respectively). Mean probability for same‐day discharge in scheduled same‐day patients was significantly increased in discharged vs. admitted patients (69.90% SD: 7.6 vs. 67.03 SD: 8.3 p < 0.001), but not in admitted same‐day versus planned in‐patients (67.52% SD: 7.5 p = 0.14). The potential increase in same‐day discharge at the current same‐day surgical capacity of 82.9% of procedures was 2.2% but increased with lower capacities.
Conclusions
Machine learning based prognostic probability scores for planning same‐day hip and knee replacement in pre‐selected eligible patients did not provide relevant potential increases in same‐day discharge rates.
Editorial Comment
This study assessed if an advanced model using routinely available clinical data could confidently predict whether of not cases planned for same‐day hip or knee arthroplasty would be successfully discharged as planned. Data from multiple collaborating fast‐track surgical centers in Denmark contributed to the model. The advanced model here based on the available clinical data did not perform clearly better than other simpler predictive models that have already been reported.