Assessment of the performance of classifiers in the discrimination of healthy adults and elderly individuals through functional fitness tasks

Purpose The identification of features that aid in the assessment of processes related to aging is an area of ever-increasing study. The use of biomarkers and tools that offer greater specificity and sensitivity to quantify and characterize motor tasks, aid in the understanding of biological changes that occur due to advancing age, and as such predefine phenotypes related to age or health outcomes. The proposal behind this study is thus to assess the performance of different classifiers in the discrimination of both healthy individual adults and senior citizens from the characterization of functional fitness tasks, using inertial sensors. Method Ninety-nine healthy participants were recruited, with ages ranging from 20 to 98 years old. The collection of data was performed by means of two inertial measurement units (IMUs), positioned in the region of the distal third of the forearm of the dominant hand and on the back of the dominant hand. The participants performed three successive tasks with the forearm flexed, (i) at rest, (ii) pulp to pulp pinch, and (iii) supination/pronation. Different features were extracted from the signals that were then used in the comparison of the values for specificity, sensitivity, precision, and accuracy. The classifiers random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN), and naive Bayes (NB) were used to classify the groups. Results Through use of the features extracted by the IMUs in the region of the distal third of the forearm of the dominant hand, and on the back of the dominant hand of the volunteers, the classifier that presented the best sensitivity was the SVM, when fed with 25% of the features, with a rate of 89.6%. The RF classifier was the one that obtained the best specificity (72.8%), when fed with all the features. However, the NB obtained the best precision and accuracy (75.5% and 79.3% respectively), when fed with 60% of the features. Conclusion The results obtained in this study demonstrate that the use of classification algorithms from machine learning in the discrimination of healthy adult and senior citizen groups, with high rates of sensitivity and specificity, provide valuable information for clinical assessment concerning the prediction of motor changes related to advancing age regarding the characterization of motor-related tasks.