Comprehensive Summary
This study developed and validated an artificial intelligence (AI) platform to predict prolonged dependence on mechanical ventilation in patients with critical orthopedic trauma. A total of 1,400 patients from the Medical Information Mart for Intensive Care III (MIMIC-III) database were analyzed, with prolonged mechanical ventilation defined as an inability to wean from ventilation for seven or more days. The median patient age was 55 years (IQR: 36–73), with 65.0% male and 35.0% female. Patients were randomly divided into training (80%) and validation (20%) cohorts, and six machine learning models were evaluated: logistic regression (LR), extreme gradient boosting machine (eXGBM), decision tree (DT), random forest (RF), support vector machine (SVM), and light gradient boosting machine (LightGBM). The eXGBM model demonstrated the highest predictive performance, scoring 50 out of 60 in evaluation metrics. It achieved an area under the curve (AUC) of 0.949 (95% CI: 0.933–0.961), an accuracy of 87.1%, a recall of 0.892, a Brier score of 0.088, and a calibration slope of 0.999, indicating excellent reliability. The LightGBM model followed closely with an AUC of 0.953 and an accuracy of 89.0%, while the RF model demonstrated strong discrimination with an AUC of 0.936. External validation in 122 critically ill orthopedic trauma patients confirmed strong model performance, yielding an AUC of 0.893 (95% CI: 0.819–0.967). The AI model was deployed as a risk calculator via an online platform, allowing clinicians to input patient data and receive automated risk assessments for ventilation dependency. Feature importance analysis using SHapley Additive exPlanations (SHAP) identified respiratory rate, lower limb fracture, glucose levels, PaO2, and PaCO2 as the five most significant predictive variables. Patients with spine fractures had a 62.6% higher risk of prolonged ventilation, while patients with lower limb fractures had a lower risk (p = 0.025). These results highlight the effectiveness of machine learning in stratifying ICU patients and optimizing mechanical ventilation management strategies.
Outcomes and Implications
This AI-driven model offers a groundbreaking tool for early identification of ICU patients at risk of prolonged mechanical ventilation, a critical factor affecting morbidity, healthcare costs, and ICU resource allocation. Given that 15.64% of ICU patients with orthopedic trauma in this study required prolonged ventilation, accurate risk stratification is crucial. The AI model's predictions allow for personalized intervention strategies, potentially reducing ventilator-associated pneumonia, muscle atrophy, and increased mortality risk. The study identified that higher glucose levels (p < 0.001), lower PaO2 (p = 0.006), and elevated PaCO2 (p < 0.001) were significantly associated with prolonged ventilation dependency. Additionally, patients with spine fractures had a 62.6% increased risk of requiring long-term mechanical ventilation compared to those with lower limb fractures. By integrating machine learning into ICU workflows, this model could significantly enhance risk stratification, leading to early interventions such as intensified respiratory therapy, optimized sedation strategies, and timely tracheostomy evaluation. The AI tool provides real-time, automated predictions and clinical decision support, allowing for hospital resource optimization, reduced ICU stay durations, and improved patient outcomes. Patients classified as high risk may benefit from aggressive ventilation weaning protocols, while low-risk patients could transition to non-invasive respiratory support earlier, potentially lowering ICU costs. Future research should explore multicenter external validation and refine models by incorporating biochemical markers, imaging data, and genomic risk factors. Expanding the AI tool to include continuous real-time monitoring could further enhance predictive capabilities and improve patient outcomes in critical care settings. The successful deployment of this AI-driven platform marks a significant step toward data-driven decision-making in critical care medicine, enabling personalized, precise, and proactive patient management.