Eunsun Oh1,2, Sung Wook Seo3, Young Cheol Yoon1, Dong Wook Kim3, Sunyoung Kwon4, Sungroh Yoon4. 1. 1 Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea. 2. 2 Department of Radiology, Soonchunhyang University Seoul Hospital, Seoul, Korea. 3. 3 Department of Orthopedic Surgery, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea. 4. 4 Department of Electrical and Computer Engineering and Interdisciplinary Program in Bioinformatics, Seoul National University, Seoul, Korea.
Abstract
PURPOSE: The purpose of this article is to compare the predictive power of two models trained with computed tomography (CT)-based radiological features and both CT-based radiological and clinical features for pathologic femoral fractures in patients with lung cancer using machine learning algorithms. METHODS: Between January 2010 and December 2014, 315 lung cancer patients with metastasis to the femur were included. Among them, 84 patients who underwent CT scan and were followed up for more than 3 months were enrolled. We examined clinical and radiological risk factors affecting pathologic fracture through logistic regression. Predictive analysis was performed using five different supervised learning algorithms. The power of predictive model trained with CT-based radiological features was compared to those trained with both CT-based radiological and clinical features. RESULTS: In multivariate logistic regression, female sex (odds ratio = 0.25, p = 0.0126), osteolysis (odds ratio = 7.62, p = 0.0239), and absence of radiation therapy (odds ratio = 10.25, p = 0.0258) significantly increased the risk of pathologic fracture in proximal femur. The predictive model trained with both CT-based radiological and clinical features showed the highest area under the receiver operating characteristic curve (0.80 ± 0.14, p < 0.0001) through gradient boosting algorithm. CONCLUSION: We believe that machine learning algorithms may be useful in the prediction of pathologic femoral fracture, which are multifactorial problem.
PURPOSE: The purpose of this article is to compare the predictive power of two models trained with computed tomography (CT)-based radiological features and both CT-based radiological and clinical features for pathologic femoral fractures in patients with lung cancer using machine learning algorithms. METHODS: Between January 2010 and December 2014, 315 lung cancerpatients with metastasis to the femur were included. Among them, 84 patients who underwent CT scan and were followed up for more than 3 months were enrolled. We examined clinical and radiological risk factors affecting pathologic fracture through logistic regression. Predictive analysis was performed using five different supervised learning algorithms. The power of predictive model trained with CT-based radiological features was compared to those trained with both CT-based radiological and clinical features. RESULTS: In multivariate logistic regression, female sex (odds ratio = 0.25, p = 0.0126), osteolysis (odds ratio = 7.62, p = 0.0239), and absence of radiation therapy (odds ratio = 10.25, p = 0.0258) significantly increased the risk of pathologic fracture in proximal femur. The predictive model trained with both CT-based radiological and clinical features showed the highest area under the receiver operating characteristic curve (0.80 ± 0.14, p < 0.0001) through gradient boosting algorithm. CONCLUSION: We believe that machine learning algorithms may be useful in the prediction of pathologic femoral fracture, which are multifactorial problem.