Comprehensive Summary
This study by Chen et al. uses a multiclass machine learning model to predict length of stay of patients following primary total knee arthroplasty (TKA). Data gathered from over 150,000 patients over the age of 18 years old was used to train and test the model to classify patients into three groups including same-day discharge, 1-3 day length of stay (LOS), and prolonged LOS (>3 days). Four different models– artificial neural network (ANN), random forest (RF), k-nearest neighbor (KNN), and XGBoost– were tested with their various levels of accuracy recorded. The study found that the random forest model outperformed all the others with an accuracy of 90.3%. The strongest predicting variables to LOS included anesthesia type, sex, BMI, American Society of Anesthesiologists score, hypertension, age, and operation time. Previous machine models have been developed to attempt to classify LOS following total joint arthroplasties; however, they have mainly been limited to binary classifications which proves more difficult to accurately predict LOS in more narrow ranges. The authors were able to develop one of the first multiclass learning models attempting to predict LOS for patients following TKA with high accuracy.
Outcomes and Implications
The number of TKAs has been rising in the United States recently, placing strains on hospitals as one of the most commonly performed orthopedic surgeries. Being able to accurately predict LOS for patients following TKA would allow hospitals to reduce costs by being able to better plan resource allocation while also informing physicians on what patients may be more suited for out-patient TKA procedures. This work has clear clinical implications with its predictive power in determining patient LOS following TKA; however, no clear timeline was given for clinical implementation. A few limitations, namely more external validation studies, persist which may limit the current clinical applicability of the model. Overall, the model developed by Chen et al. has high potential for its application in assisting in stronger predictions of LOS for patients following TKA.