Can Classification and Regression Tree Analysis Help Identify Clinically Meaningful Risk Groups for Hip Fracture Prediction in Older American Men (The MrOS Cohort Study)?

Although the WHO fracture risk algorithm (FRAX) is used to predict fracture, the utility of some simple machine-learning methods, such as classification and regression trees (CARTs) should be evaluated to determine their efficacy in fracture prediction. Follow-up time for the hip fracture analyses of 5977 community-dwelling American men aged ≥65 years old was truncated to 10 years. There were 172 (2.9%) men who had an incident nontraumatic hip fracture. The CARTs were developed using hip BMD and common clinical risk factors as follows: model 1 = using classification with continuous variables of age, total hip BMD, and femoral neck BMD, or together with common clinical risk factors; and model 2 = using classification with continuous variables of age, total hip BMD, femoral neck BMD, FRAX score, osteoporosis by -score at the hip, and common clinical risk factors. The predictive performance of risk models derived from CARTs was compared with the basic classification of FRAX at 3% (basic model). From model 1, discriminators selected by CART were total hip BMD, age, and femoral neck BMD; no other clinical risk factors were selected. From model 2, discriminators selected by CART were FRAX score, femoral neck BMD, and age. Compared with the basic model using only a high-risk group by FRAX ≥3%, no significantly improved predictive performance was demonstrated by model 1 or model 2 as identified by CART with the area under the receiver-operating characteristic curve for each model of 0.714 (95% CI, 0.676 to 0.751) or 0.726 (95% CI, 0.690 to 0.762) versus 0.703 (95% CI, 0.667 to 0.740), respectively. The improved overall net reclassification improvement index was 0.02 (95% CI, -0.04 to 0.08) and 0.05 (95% CI, -0.01 to 0.10), respectively. Although a FRAX category is a good clinical indicator for hip fracture risk, a simple classification by age and BMD may provide an alternative way to estimate a clinical risk level of 3.0%. © 2019 The Authors. is published by Wiley Periodicals, Inc. on behalf of the American Society for Bone and Mineral Research.