Machine Learning-Based Diagnostic Model for Preoperative Anxiety in Elderly Patients with Hip Fractures and SHAP Analysis

Technical Summary: Machine Learning-Based Diagnostic Model for Preoperative Anxiety in Elderly Patients with Hip Fractures

Problem Statement
Preoperative anxiety significantly impacts the treatment prognosis of elderly patients with hip fractures, leading to adverse outcomes such as difficult anesthetic management, suboptimal pain control, delayed recovery, and increased complications. However, diagnosing and screening for anxiety in this specific demographic is challenging. Clinical manifestations of anxiety (e.g., palpitations, restlessness) often overlap with acute post-fracture pain, stress responses, and age-related cognitive or communication difficulties. Furthermore, the typically lower educational attainment of elderly patients complicates the use of standard diagnostic tools, often resulting in missed opportunities for timely psychological intervention. Currently, there is a scarcity of studies establishing specific diagnostic models for preoperative anxiety in this population.

Methodology
This retrospective study was conducted at the Third Hospital of Hebei Medical University, enrolling elderly patients (aged ≥65) with femoral neck or intertrochanteric fractures admitted between November 2020 and May 2025.

• Data Collection: The study initially collected 74 potential predictive features, including demographics, fracture characteristics, comorbidities (e.g., hypertension, coronary artery disease, heart failure), and extensive laboratory parameters (admission values and min/max values during hospitalization).
• Outcome Definition: Anxiety was defined using the Generalized Anxiety Disorder-7 (GAD-7) questionnaire administered the day prior to surgery. A score >4 classified patients into the "anxiety group," while a score ≤4 formed the "non-anxiety group."
• Feature Selection: From the initial 74 features, a combination of the Least Absolute Shrinkage and Selection Operator (LASSO) regression and the Boruta algorithm was employed to identify the most significant predictors. This process reduced the feature set to nine core variables.
• Model Development: The cohort of 400 patients (220 with anxiety, 180 without) was split into a training set (70%) and a test set (30%). Seven machine learning algorithms were trained and tuned using Bayesian optimization with five-fold cross-validation: Logistic Regression (LR), Neural Network (NN), Support Vector Machine (SVM), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost).
• Evaluation and Interpretation: Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), accuracy, precision, recall, F1 score, specificity, calibration curves (Brier score), and Decision Curve Analysis (DCA). The optimal model was further interpreted using SHAP (SHapley Additive exPlanations) analysis to quantify feature importance and directionality.

Key Contributions

• Diagnostic Model Construction: The study successfully developed and compared seven distinct machine learning diagnostic models specifically for preoperative anxiety in elderly hip fracture patients.
• Feature Identification: Through rigorous feature selection, the study identified nine core clinical predictors: Visual Analog Scale (VAS) pain score, age, hypertension (HTN), coronary artery disease (CAD), minimum hemoglobin (Hb Min), marital status, acute heart failure (AHF), Charlson Comorbidity Index (CCI), and chronic heart failure (CHF).
• Interpretability: Unlike "black-box" models, the study applied SHAP analysis to the best-performing model, providing a transparent visualization of how specific clinical variables influence the prediction of anxiety risk.

Results

• Feature Selection: The intersection of LASSO and Boruta algorithms yielded nine non-collinear features.
• Model Performance: Among the seven algorithms, the CatBoost model demonstrated superior performance.
  • Training Set: Accuracy of 0.861, Precision of 0.883, Recall of 0.853, F1 score of 0.868, and Specificity of 0.869.
  • Test Set: Accuracy of 0.775, Precision of 0.841, Recall of 0.757, F1 score of 0.797, and Specificity of 0.800.
  • Comparative Metrics: CatBoost achieved the highest AUC of 0.827 in the test set, significantly outperforming LR, NN, RF, XGBoost, and LightGBM (p < 0.05). It also exhibited the best calibration (Brier score: 0.1714) and the most favorable decision curve analysis.
• SHAP Analysis Findings:
  • Positive Correlations (Increased Risk): Higher VAS pain scores, older age, and higher CCI scores were positively correlated with anxiety risk. The presence of hypertension, coronary artery disease, acute/chronic heart failure, and widowed marital status also increased risk.
  • Negative Correlation (Decreased Risk): Lower minimum hemoglobin (Hb Min) levels were negatively correlated with anxiety risk in the model's output (though the discussion notes anemia generally exacerbates physiological stress, the specific SHAP direction indicated a negative association in this dataset context).

Significance and Claims
The authors claim that this study provides a valuable, objective tool for clinicians to proactively screen high-risk elderly patients for preoperative anxiety. By identifying specific risk factors—particularly pain levels, cardiovascular comorbidities, and social status—the CatBoost model enables the formulation of personalized perioperative management plans. The study suggests that early identification of these patients can facilitate timely psychological interventions, tailored anesthesia, and optimized care, ultimately improving surgical experiences and recovery outcomes. The authors emphasize that while the model shows promise, it is a single-center study with a relatively small sample size and observational design, limiting causal inference and generalizability, thus requiring further validation in larger, diverse populations.