Petri net transition times as training features for multiclass models to support the detection of neurodegenerative diseases

This paper proposes the transition times of Petri net models of human gait as training features for multiclass random forests (RFs) and classification trees (CTs). These models are designed to support screening for neurodegenerative diseases. The proposed Petri net describes gait in terms of nine cyclic phases and the timing of the nine events that mark the transition between phases. Since the transition times between strides vary, each is represented as a random variable characterized by its mean and standard deviation. These transition times are calculated using the PhysioNet database of vertical ground reaction forces (VGRFs) generated by feet-ground contact. This database comprises the VGRFs of four groups: amyotrophic lateral sclerosis, the control group, Huntington's disease, and Parkinson disease. The RF produced an overall classification accuracy of 91%, and the specificities and sensitivities for each class were between 80% and 100%. However, despite this high performance, the RF-generated models demonstrated lack of interpretability prompted the training of a CT using identical features. The obtained tree comprised only four features and required a maximum of three comparisons. However, this simplification dramatically reduced the overall accuracy from 90.6% to 62.3%. The proposed set features were compared with those included in PhysioNet database of VGRFs. In terms of both the RF and CT, more accurate models were established using our features than those of the PhysioNet.