gms | German Medical Science

Objectives: In-hospital resource utilization and length of stay (LOS) are increased in obese patients undergoing total joint arthroplasty. Such factors are disincentives for institutions participating in bundle-payment programs. Given the worsening obesity epidemic and increased utilization of arthroplasty in this population, we investigated whether prolonged hospitalization in obese patients undergoing total hip and total knee arthroplasty (THA and TKA) could be predicted based on pre-operative variables.

Methods: The arthroplasty registry of a single academic center included 4,563 obese patients who underwent unilateral THA or TKA between 2020 and 2021. No data was imputed. A total of 31 pre-surgical parameters were included as predictor variables. The data corpus was partitioned into training (80%) and hold-out test (20%). Both binary classification and regression predictive modeling approaches were undertaken, training 7 distinct machine learning models. For binary modelling, LOS was split into patients staying less than 2 nights (41.36%) and patients staying 2 nights or more (58.46%). Binary classification model performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC), Matthews correlation coefficient (MCC), F1 score, accuracy, sensitivity, precision, and the Brier score. Regression models were evaluated utilizing Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). 95% confidence intervals (95% CI) were calculated via bootstrapping.

Results and conclusion: The explainable boosted machine (EBM) model was the best performing binary model: F1-Score 0.75 (95% CI 0.72, 0.78); accuracy 0.69 (0.67, 0.72); sensitivity 0.71 (0.67, 0.75); precision 0.74 (0.7, 0.78), AUC-ROC 0.75 (0.72, 0.79), MCC 0.38 (0.32, 0.44) and Brier Score 0.1990. The quantile regression forest (QRF) was the best performing regression model with MAE of 21.49 (20.01, 23.23) and RMSE of 32.31 (28.82, 39.97) hours. Male sex, older age, and not being married were the most important predictors for increased length of stay across both models.
Correlation of QRF predicted LOS and actual LOS revealed a Spearman Correlation Coefficient of 0.43 (p <0.001). All models demonstrated robust predictive capacity emphasizing their potential clinical utility.

The developed machine learning models facilitate enhanced comprehension of factors linked to divergent hospital LOS in obese patients undergoing THA or TKA, exhibiting that unmodifiable demographic factors are the strongest predictors of LOS in this population.

Figure 1 [Fig. 1]