Table 5 Area under the curve of diagnostic factor. Mantzaris et al. 2009 Diagnostic factor Age Gender Height AUC 0.646 0.503 0.560 Weight 0.641 Page 9 of 17 Cruz et al. BioMed Eng OnLine (2018) 17:12 combinations of parameters, in order to identify the predictive value of each, the best combination of parameters was: age, BMC, BMD and the Hurst Fractal exponent.

A more comprehensive study was conducted in 2013 in a group of 1674 medical records of Korean postmenopausal women, whose goal was to assess the risk of osteoporosis. In this study a comparison was made with several models based on popular machine learning algorithms such as SVM, RF, ANN and LR with four conventional clinical decision instruments: OST, ORAI, SCORE and OSIRIS (see Table 6). The quantitative of 1674 women had osteoporosis at any of the following locations: hip, femoral neck or lumbar spine. The training set was 1000 patients, the remaining (674 patients) were used as a test to predict osteoporosis in postmenopausal women. The SVM predicted risk of osteoporosis with an AUC of 0.827, accuracy of 76.7%, sensitivity of 77.8% and specificity of 76.0% in the total hip, femoral neck or lumbar spine. The somatic factors selected by SVM were age, height, weight, body mass index, duration of menopause, duration of breastfeeding, estrogen therapy, hyperlipidemia, hyper

In this study SVM presented the best results under the receiver operating characteristic curve (ROC) of ANN, LR, OST, ORAI, SCORE and OSIRIS for the training set.

Still in 2013, Sapthagirivasan and Anburajan have proposed the use of the trabecular border on digital hip radiographs for the identification of osteoporosis. Applying a kernel-based SVM, they were able to extract trabecular features from digital hip radiographs, identifying individuals with low BMD. The group involved in this study was 50 women from southern India with no previous history of osteoporotic fracture, of these 28 were used for training and 22 for tests. The best results were achieved with a quintuple cross validation analysis with mean accuracy of 90%. It is important to note that in Table 6 Selection of variables in machine learning and conventional methods for osteoporosis risk of hip, neck and lumbar Variables Machine learning method Conventional method SVM RF ANN LR OST ORAI SCORE Age o o o o o o o Height o o o Weight o o o o o o o Body mass index o o o Waist circumference o Pregnancy o o Duration of menopause o