Classification Algorithms for Hip Fracture Prediction Based on Recursive Partitioning Methods

Abstract

This article presents 2 modifications to the classification and regression tree. The authors improved the robustness of a split in the test sample approach and developed a cost-saving classification algorithm by selecting noninferior to the optimum splits from variables with lower cost or being used in parent splits. The new algorithm was illustrated by 43 predictive variables for 5-year hip fracture previously documented in the Study of Osteoporotic Fractures. The authors generated the robust optimum classification rule without consideration of classification variable costs and then generated an alternative cost-saving rule with equivalent diagnostic utility. A 6-fold cross-validation study proved that the cost-saving alternative classification is statistically noninferior to the optimal one. Their modified classification and regression tree algorithm can be useful in clinical applications. A dual X-ray absorptiometry hip scan and information from clinical examinations can identify subjects with elevated 5-year hip fracture risk without loss of efficiency to more costly and complicated algorithms.