Optimizing Sensitivity in Machine Learning Models for Pediatric Post-operative Kyphosis Prediction

Post-operative kyphosis represents a significant complication following pediatric spinal corrective surgery, necessitating sophisticated prediction methods to identify high-risk patients. This study developed and evaluated machine learning models for kyphosis prediction using a dataset of 81 pediatric patients by comparing the logistic regression and decision tree approaches. Despite achieving a higher overall accuracy (82%), the logistic regression model failed to identify any kyphosis cases, rendering it clinically ineffective. Conversely, the decision tree model demonstrated superior clinical utility by successfully identifying 33% of kyphosis cases while maintaining 71% accuracy. Feature importance analysis established starting vertebral position as the dominant predictor (importance=0.554), followed by patient age (0.416), with vertebrae count contributing minimally (0.030). The decision tree identified critical thresholds for risk stratification: operations beginning at or above T8-T9, particularly in children aged 5-9 years, carried a substantially elevated kyphosis risk. Our methodological approach emphasizes sensitivity over conventional accuracy metrics, recognizing that missing high-risk patients have greater clinical consequences than unnecessary monitoring. This study demonstrates the capacity of decision tree models to extract clinically meaningful patterns from small, imbalanced surgical datasets that elude conventional statistical approaches.