Poster Presentation ESA-SRB-APEG-NZSE 2022

Prediction models for new fragility fracture using penalized regression and artificial intelligence in Korean women aged 50 years or older and men aged 60 years or older (#284)

Sihoon Lee 1 , Jung-Hyun Kim 2 , Soo-ryun Lee 2 , Jinseub Hwang 3 , Da Young Kang 3 , Ga Eun Lee 3 , Yoon-Sok (Martin) Chung 2
  1. Gachon University, Incheon, South Korea
  2. Ajou University, Suwon, South Korea
  3. Daegu University, Gyeongsan, South Korea

The osteoporotic fracture could be predicted by clinical risk factors, bone mineral density (BMD), and bone turnover marker (BTM). This study was conducted on 6 University Hospitals in Korea with subjects of women aged 50 years or older and men aged 60 years or older retrospectively after approval by the Institutional Review Boards. We developed prediction models for new fractures based on age, sex, height, weight, previous fracture, current smoking, glucocorticoids, rheumatoid arthritis, secondary osteoporosis, femoral neck BMD, lumbar spine (L1~4) BMD, and total alkaline phosphatase (TALP). Fracture Risk Assessment Tool (FRAX) scores with the World Health Organization had been calculated as Korean. Among the first collected 28,508 subjects, 18,708 participants had been included after applying the exclusion criteria for the purpose of this study. A total of 971 new fractures occurred during the 1.35 years of mean duration of follow-up. The whole dataset was randomly split into training and test sets in a 7:3 ratio. In this study, we applied 3 penalized regression models (Lasso, Ridge, Elastic-Net), 2 machine-learning models (random forest; RF, extreme gradient boosting machine; XGBoost), and 1 deep-learning model (deep neural network; DNN). To evaluate the performance of models, we used accuracy and area under the receiver operating characteristics (AUROC) curve. The accuracy in test sets was 94.50% with a cut-off of 0.5 for all models. The AUROC in test sets were 0.734 for Lasso, 0.736 for Ridge, 0.734 for Elastic-Net, 0.544 for RF, 0.737 for XGBoost, and 0.732 for DNN.