Study of Hip Fracture Risk using Tree Structured Survival Analysis

In this paper we studied the risk of hip fracture for post-menopausal women by classifying women into different subgroups based on their risk of hip fracture. The subgroups were generated such that all the women in a particular subgroup had relatively similar risk while women belonging to two different subgroups had rather different risks of hip fracture. We used the Tree Structured Survival Analysis (TSSA) method to generate the subgroups based upon the cross-sectional data from 7,665 women enrolled in the Study of Osteoporotic Fractures (SOF). All of these women had forearm, os calcis, hip and spine bone mineral density (BMD) measurements. Time to hip fracture since BMD measurement was also recorded for these women and was treated as the outcome variable. A random sample consisting of 75 % (training data set) of women from the 7,665 available was used to generate the prognostic subgroups while the other 25 % (validation data set) was used to validate the results. Based on the training data set, TSSA identified four subgroups for whom the risk of hip fracture for an average of 6.5 years of follow-up was 19 %, 9 %, 4 % and 1%. The rules to generate the subgroups were based on BMD of Ward’s triangle, BMD of the os calcis, and BMD of the femoral neck, and age. We reproduced these results using the validation data set, showing the usefulness of the classification rules in a clinical setting. In conclusion, TSSA provided a useful, powerful and reproducible procedure for identification of meaningful prognostic subgroups based upon an individual woman’s age and BMD measurements. J Miner Stoffwechs 2003; 10: 11– 16.