Construction of a machine learning-based prediction model for intraoperative acquired pressure injury in children undergoing neurosurgery

Technical Summary: Construction of a Machine Learning-Based Prediction Model for Intraoperative Acquired Pressure Injury in Children Undergoing Neurosurgery

Problem Statement
Intraoperative acquired pressure injury (IAPI) is a frequent and serious complication among children undergoing neurosurgery. Due to the delicate nature of these procedures, the necessity for forced positioning, the use of instruments causing local tissue compression, and prolonged postoperative bed rest, pediatric patients face elevated risks. While previous studies have identified an incidence rate of approximately 11.5% in neurosurgical children, the specific risk factors for this population remain incompletely understood. Furthermore, existing prediction tools, such as a previously developed nomogram, demonstrated only moderate predictive efficacy (c-index = 0.79). To date, no study has reported the use of machine learning (ML) techniques to construct a prediction model for IAPI in neurosurgical children.

Methodology
This retrospective study utilized clinical data from 776 children (aged ≤18 years) who underwent neurosurgery at a tertiary children's hospital in Chongqing between January and June 2023. The study adhered to a 10 events per variable (EPV) principle for sample size calculation, ultimately including 14 potential risk factors.

• Data Collection & Preprocessing: Data included demographic information, preoperative laboratory values (hemoglobin, hematocrit, BMI, Braden Q score), and intraoperative variables (ASA classification, surgical position, operation/anesthesia duration, blood loss, use of power tools, and hypothermia). IAPI was diagnosed according to 2019 NPUAP standards.
• Feature Selection: The Least Absolute Shrinkage and Selection Operator (LASSO) regression was employed to screen variables and reduce dimensionality, selecting 13 features with non-zero coefficients.
• Model Construction: The dataset was split into a training set (70%) and a validation set (30%). Nine machine learning algorithms were implemented: Logistic Regression (LR), Random Forest (RF), Neural Network (NNET), Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), Naive Bayes (NB), K-Nearest Neighbors (KNN), and Decision Tree (DT).
• Optimization & Validation: Models were tuned using ten-fold cross-validation and grid search on the training set. Performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, precision, F1 score, and Decision Curve Analysis (DCA).
• Interpretability: The optimal model was analyzed using SHapley Additive exPlanations (SHAP) to visualize feature importance and explain the decision-making process.

Key Results

• Incidence: The observed incidence of IAPI in the cohort was 7.9% (62/776). The majority of injuries were Stage I (95.16%), with the cheek, forehead, and temple being the most common sites.
• Feature Selection: LASSO regression identified 13 key predictors, including age, sex, Braden Q score, preoperative hematocrit, intraoperative blood loss, operation time, anesthesia time, surgical position, use of power tools, ASA classification, surgical type, and BMI.
• Model Performance: Among the nine algorithms, the XGBoost model demonstrated the most stable and optimal performance.
  • Validation Set Metrics: XGBoost achieved an AUC of 0.7325, with a sensitivity of 0.8889, specificity of 0.4813, precision of 0.126, and an F1 score of 0.2207.
  • Comparison: While the XGBoost model showed slightly lower AUC compared to the authors' previous nomogram model (0.79), it demonstrated higher sensitivity. Decision Curve Analysis indicated that XGBoost provided the greatest net clinical benefit in the validation set.
• Risk Factor Analysis: SHAP visualization identified the top six risk factors driving the model's predictions: intraoperative blood loss, surgical position, patient age, operation duration, anesthesia duration, and the use of external force (e.g., drills/milling cutters).

Key Contributions

1. Novel ML Application: This study constructs a machine learning-based prediction model for IAPI in children undergoing neurosurgery, addressing a gap where no such model had previously been reported.
2. Algorithmic Comparison: It systematically compares nine distinct machine learning algorithms, identifying XGBoost as the superior model for this specific clinical context.
3. Interpretability: By integrating SHAP analysis, the study moves beyond "black box" predictions, providing a transparent visualization of how specific clinical indicators (such as blood loss and surgical position) influence the risk of injury.
4. Refined Risk Factors: The study re-evaluates risk factors, confirming the significance of intraoperative blood loss and the use of power tools, which were not as prominently highlighted in previous non-ML models.

Significance and Claims
The authors claim that the constructed XGBoost-based prediction model possesses "certain predictive efficacy." The primary significance of this work lies in its potential to assist clinical nurses in the early identification of high-risk pediatric neurosurgical patients. By leveraging this model, healthcare providers can take personalized preventive measures earlier in the care continuum. The study emphasizes that while the model is not a perfect replacement for clinical judgment, it offers a scientific basis for screening and helps streamline the assessment process by highlighting the most critical risk factors, thereby potentially reducing assessment time and improving the accuracy of IAPI prevention strategies. The authors remain modest regarding the model's generalizability, noting limitations such as the retrospective nature of the data, the lack of external validation, and the need to explore interactions between risk factors in future studies.