Trustworthy AI models to predict prolonged hospital length-of-stay in orthopedic surgery

Figures

• 

  Fig. 1. Calibration performance of the Optuna-optimized XGBoost model based on 85% OOF cross-validation predictions. a, b Calibration curves for prolonged LOS defined as >5 and >2 days. c, d The corresponding Brier scores across evaluated models.

• 

  Fig. 2. Calibration performance of the Optuna-optimized XGBoost model on the independent 15% hold-out test set. a, b Calibration curves for prolonged LOS defined as >5 and >2 days. c, d The corresponding Brier scores across evaluated models.

• 

  Fig. 3. Subgroup calibration analysis of the Optuna-optimized XGBoost model based on 85% OOF cross-validation predictions. a, b Calibration curves stratified by sex for prolonged LOS defined as >5 and >2 days, respectively. c, d The corresponding calibration curves stratified by age groups for LOS defined as >5 and >2 days, respectively. Across all subgroups, predicted probabilities demonstrated close alignment with the diagonal reference line, indicating good calibration and agreement between predicted and observed outcomes without evidence of systematic bias. Slightly greater variability was observed in certain age strata at intermediate probability levels for LOS > 2 days. With that, calibration remained robust across clinically relevant thresholds, with marginally more consistent performance observed for LOS > 5 days. Age groups were encoded as: Group 1 (≤50), Group 2 (51–64), Group 3 (65–79), and Group 4 (≥80 years). Sex groups were encoded as: Group 0 (female) and Group 1 (male).

• 

  Fig. 4. Subgroup calibration analysis of the Optuna-optimized XGBoost model on the independent 15% hold-out test set. a, b Calibration curves stratified by sex for prolonged LOS defined as >5 and >2 days, respectively. c, d The corresponding calibration curves stratified by age groups for LOS defined as >5 and >2 days, respectively. Across all subgroups, predicted probabilities demonstrated close alignment with the diagonal reference line, indicating good calibration and agreement between predicted and observed outcomes without evidence of systematic bias. Slightly greater variability was observed in certain age strata at intermediate probability levels for LOS > 2 days. Overall, calibration remained robust across clinically relevant thresholds, with marginally more consistent performance observed for LOS > 5 days. Age groups were encoded as: Group 1 (≤50), Group 2 (51–64), Group 3 (65–79), and Group 4 (≥80 years). Sex groups were encoded as: Group 0 (female) and Group 1 (male).

• 

  Fig. 5. Fairness assessment of the Optuna-optimized XGBoost model based on 85% OOF cross-validation for predicting prolonged LOS, using equal opportunity, reported as subgroup-specific TPR. Panels show TPRs stratified by: (A) age group, (B) sex, and (C) race. Age groups were encoded as: Group 1 (≤50), Group 2 (51–64), Group 3 (65–79), and Group 4 (≥80 years). Sex groups were encoded as: Group 0 (female), Group 1 (male), and Group 2 (non-binary/other). Race groups were encoded following NSQIP conventions as: Group 1 (White), Group 2 (Black or African American), Group 3 (American Indian or Alaska Native), Group 4 (Native Hawaiian or Other Pacific Islander), Group 5 (Asian), and Group 6 (Unknown or not reported). a Prolonged LOS defined as >5 days. b Prolonged LOS defined as >2 days.

• 

  Fig. 6. Fairness assessment of the Optuna-optimized XGBoost model on the independent 15% hold-out test set for predicting prolonged LOS, using equal opportunity, reported as subgroup-specific TPR. Panels show TPRs stratified by: (A) age group, (B) sex, and (C) race. Age groups were encoded as: Group 1 (≤50), Group 2 (51–64), Group 3 (65–79), and Group 4 (≥80 years). Sex groups were encoded as: Group 0 (female), Group 1 (male), and Group 2 (non-binary/other). Race groups were encoded following NSQIP conventions as: Group 1 (White), Group 2 (Black or African American), Group 3 (American Indian or Alaska Native), Group 4 (Native Hawaiian or Other Pacific Islander), Group 5 (Asian), and Group 6 (Unknown or not reported). a Prolonged LOS defined as >5 days. b Prolonged LOS defined as >2 days.

• 

  Fig. 7. Global and local explainability of the Optuna-optimized XGBoost model for prolonged LOS prediction. The top row, (a, b), shows SHAP-based global explanations for the LOS > 5 days and LOS > 2 days outcome definitions, respectively. The second and last rows, (c, d), show LIME-based local explanations for the same outcome definitions. For binary classification, Class 0 corresponded to LOS ≤ 5 days (or LOS ≤ 2 days in the alternative definition), whereas Class 1 corresponded to prolonged LOS.

• 

  Fig. 8. Clinician-facing decision-support application interface illustrating model predictions and explainability at both population and individual levels. a Global SHAP summary of feature importance at the population level. b Local SHAP explanation for an individual patient. c LIME-based local explanation for an individual patient.

• 

  Fig. 9. Study cohort derivation and demographic distribution stratified by LOS. The cohort includes orthopedic inpatient surgical cases from the NSQIP dataset (2021–2023), stratified into Class 0 (LOS≤5 days) and Class 1 (LOS > 5 days) using LOS defined as >5 days. Also, Class 0 (LOS≤2 days) and Class 1 (LOS > 2 days) employing LOS defined as >2 days. Age group distributions and sex composition within each age group are shown for both LOS categories.