In addition to the advantage of being non-invasive, the design of diagnostic systems based on medical data processing has a special place in increasing knowledge, understanding, and clinical understanding. Cognitive systems in the field of Parkinson's disease first require an accurate diagnosis of the severity of tremor in patients. Therefore, this study aims to identify a topological pattern of tremor time series in Parkinson's patients and tries to distinguish between two severe and mild classes of the level of this disorder. The desired hypothesis is that the topology of points reconstructed from tremor time series in phase space contains richer information for the analysis of non-Gaussian, non-linear and non-stationary time series. Given the lack of analytical tools for extracting such information, in this study, a new processing approach based on following the basic states of the reconstructed phase trajectory of tremor time series based on compass directions was presented. Then, the behavioral pattern related to tremor severity was calculated in the form of a tactic based on dynamic feature extraction and identified in the form of a tactic based on statistical-competitive analysis. Using the KNN classifier with supervised Holdout validation, the severe and mild tremor classes were separated with 100% accuracy for the training data and 96.55% accuracy for the test data. The findings showed that a behavioral pattern consisting of north, east, and northwest directions in the tremor phase trajectory is associated with the severity of Parkinson's disease. Thus, an increase in relative variability (coefficient of variation), simultaneously with a decrease in the irregularity (Shannon entropy) of the selected directions during the time series, indicates a patient with severe tremor, and vice versa, a decrease in relative variability simultaneously with an increase in irregularity in the selected directions indicates mild tremor.