Comparing Machine Learning Methods to Improve Fall Risk Detection in Elderly with Osteoporosis from Balance Data

Figure 3

(a) ROC space of models developed by applying different machine learning techniques to unbalanced dataset (unbalanced) and balanced datasets (subsampling and oversampling), over data acquired at all conditions (open eyes, closed eyes, open eyes and closed eyes), all feature selection methods (FMSC and Weka’s methods), and all classification methods (AdaBoost, Naïve Bayes, LibSVM, Random Forest, and IBk). Models of balanced datasets using oversampling techniques have a better performance. (b) ROC space of models developed with balanced data (oversampling) over data acquired at all conditions, using all feature selection methods and all classification methods. The names of each model graphed follow this nomenclature: the first uppercase letter corresponds to set of variables used, A = all variables, F = variables selected with FSMC, and W = variables selected with Weka’s methods; the second uppercase letter corresponds to condition, Op = open eyes, Cl = closed eyes, and Me = open eyes and closed eyes; the third uppercase letter corresponds to the dataset used, O = oversampling data, and finally the rest of the name correspond to the name of the classifier used. So, A_Me_O_KNN refers to an IBk classifier built with oversampling data with the merged condition using all variables.

(a)

(b)