Signal processing, for delimitation of the target events and parametrization, is usually required when instrumented assessment is conducted to determine an individual’s functional status. However, these procedures may rule out relevant information obtained by sensors. To prevent this, the use of models based on neural networks that automatically extract relevant features from the raw signal may improve the characterization of the functional status. Thus, the aim of the study was to determine the classification accuracy of a multi-head convolutional layered neural network (CNN) using a simple functional mobility test in people with different conditions. The raw data from an inertial sensor embedded in a smartphone worn by 90 volunteers (i.e. 30 volunteers with Alzheimer’s disease, 30 with Parkinson’s disease and 30 healthy elderly people) was obtained. The CNN classification accuracy was compared to that of the two parametric classifiers, namely, linear discriminant analysis and multilayer perceptron, a neural network-based classifier. As a result, the validation process revealed that the CNN classifier correctly assigned 100% of the participants to each group. The best accuracy in pathology classification for the two parametric classifiers ranged from 55% to 88%. Therefore, the CNN model provided enhanced classification accuracy as compared to the parametric approaches, even better than the neural network-based classifier. Non parametrization may increase relevant information, thus enhancing pathology impact characterization.