Scholarlink

<p>We developed and validated a trustworthy and uncertainty-aware artificial intelligence (AI) and machine learning (ML) framework for predicting prolonged hospital length-of-stay (LOS) in orthopedic inpatient surgery using exclusively preoperative clinical and procedural data from the National Surgical Quality Improvement Program (NSQIP). LOS prediction was formulated as a binary classification task using two clinically relevant thresholds (LOS &gt; 2 days and LOS &gt; 5 days), with primary emphasis on the 5-day cutoff due to its greater clinical interpretability. Random Forest, Multilayer Perceptron, Deep Neural Network, and XGBoost models were trained using CPT-based procedure groupings and evaluated using five-fold stratified cross-validation within the 85% training cohort, with final independent evaluation performed on a completely unseen 15% hold-out test set. Model performance was assessed using accuracy, macro-averaged precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and area under the precision-recall curve (AUPRC). Across models, the 5-day threshold demonstrated the most stable and clinically meaningful performance, while results under the 2-day threshold confirmed robustness to alternative outcome definitions. The Optuna-optimized XGBoost model achieved the strongest discrimination, favorable calibration, and narrow uncertainty bounds. Algorithmic fairness was evaluated across age, sex, and race using selection rates and true positive rates, with broadly consistent subgroup performance across demographic groups. Model explainability using SHAP and LIME identified case type, CPT category, anesthesia type, and baseline physiologic status (e.g., ASA class and hematocrit) as key drivers of prolonged LOS, while patient-level explanations supported clinical plausibility. This preoperative-only and trustworthy AI framework enables transparent and equitable risk stratification to support perioperative planning, optimized bed and discharge management, and data-driven utilization management and prior authorization workflows in orthopedic care.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Farnaz Rezvani,
  • Michael R. Kann,
  • Fengyi Gao,
  • Puneet Gupta,
  • Nicole Myers,
  • Armin Kashefi,
  • Shyam Visweswaran,
  • Yanshan Wang,
  • MaCalus V. Hogan,
  • Johannes F. Plate,
  • Ahmad P. Tafti

摘要

We developed and validated a trustworthy and uncertainty-aware artificial intelligence (AI) and machine learning (ML) framework for predicting prolonged hospital length-of-stay (LOS) in orthopedic inpatient surgery using exclusively preoperative clinical and procedural data from the National Surgical Quality Improvement Program (NSQIP). LOS prediction was formulated as a binary classification task using two clinically relevant thresholds (LOS > 2 days and LOS > 5 days), with primary emphasis on the 5-day cutoff due to its greater clinical interpretability. Random Forest, Multilayer Perceptron, Deep Neural Network, and XGBoost models were trained using CPT-based procedure groupings and evaluated using five-fold stratified cross-validation within the 85% training cohort, with final independent evaluation performed on a completely unseen 15% hold-out test set. Model performance was assessed using accuracy, macro-averaged precision, recall, F1-score, area under the receiver operating characteristic curve (AUC), and area under the precision-recall curve (AUPRC). Across models, the 5-day threshold demonstrated the most stable and clinically meaningful performance, while results under the 2-day threshold confirmed robustness to alternative outcome definitions. The Optuna-optimized XGBoost model achieved the strongest discrimination, favorable calibration, and narrow uncertainty bounds. Algorithmic fairness was evaluated across age, sex, and race using selection rates and true positive rates, with broadly consistent subgroup performance across demographic groups. Model explainability using SHAP and LIME identified case type, CPT category, anesthesia type, and baseline physiologic status (e.g., ASA class and hematocrit) as key drivers of prolonged LOS, while patient-level explanations supported clinical plausibility. This preoperative-only and trustworthy AI framework enables transparent and equitable risk stratification to support perioperative planning, optimized bed and discharge management, and data-driven utilization management and prior authorization workflows in orthopedic care.