Sistema open-source para gerenciamento de dados da doença de Parkinson e detecção do sintoma do tremor em desenhos manuscritos pelo uso de histogramas de gradientes orientados

Abstract

Objective: Parkinson's disease (PD) is a neurodegenerative condition of the Central Nervous System (CNS) that causes motor and non-motor dysfunction. The disease affects 1% of the world's population over 60 years of age and remains cureless. The symptoms in CNS are chronic and progressive, affecting mainly the motor system, which brings with it the classic signs of the disease, such as tremor, stiffness, bradykinesia, and postural instability; tremor affects about 80% of patients. Thus, the monitoring of symptoms, diagnostic tests, and follow-up of the disease help in understanding and improving patient care. However, the ways of studying the disease can generate a variety of information for each patient and in large quantities (e.g., electrocardiogram examinations, electroencephalogram, imaging examinations, monitoring by inertial sensors). In this context, the motivation for this study is divided into two steps: (1) to implement a system for managing biomedical data from patients with PD, which has good usability and can be used in hospitals, clinics, research groups, and non-profit institutions that support and help people with PD; (2) to use part of this collected biomedical data focused on tremor symptom to create a low-cost tool to help decision making regarding the objective detection of the tremor in people with PD, and which uses computer vision and machine learning. Methods: For step one, the web-based system follows the Model-View-Controller (MVC) development standard, and a modularized architecture was used to provide the construction of the system by specific groups of functionalities, and to provide secure means of access and data manipulation in the system. The System Usability Scale (SUS) was used by 36 evaluators to analyze and judge the system. On the other hand, step two uses Histogram of Oriented Gradients (HOG) to describe the features of handwritten drawings (spiral and sinusoidal pattern), and together with classic Machine Learning techniques (Random Forest, K-Nearest Neighbor, and Support Vector Machine) and Deep Learning through Convolutional Neural Networks to detect the tremor symptom. At this step, hand-drawn patterns were collected from 20 participants with tremor symptoms in PD, and 20 healthy participants. Results: The multi-platform system called SIDABI was implemented in a modularized way, using responsiveness techniques to adapt the system to different devices. The SUS scale was applied and the system obtained 82.99 ± 13.97 points. The Kendall coefficient of concordance indicated 70.2% of agreement between the evaluators. In step 2, the HOG descriptor proved to be an appropriate tool in the extraction of characteristics and allowed high rates of discrimination (accuracy: 83.1%, sensitivity: 85.7%, specificity: 80.8%, area under the curve: 91%) in the set of sinusoidal patterns test using 1D Convolutional Neural Network. In addition, ANOVA and Tukey's analysis showed that the sinusoidal design is more appropriate than the spiral pattern, which is the most common design used for tremor detection. Conclusion: The use of the descriptor HOG presents a new and alternative way of quantifying and evaluating tremor by means of handwritten drawings. Moreover, it was pointed out that the sinusoidal pattern obtained better results and could be considered in clinical evaluations. Besides that, the implementation of the SIDABI system provides organized and safe data collection in a system with good usability, aiming at improving patient follow-up and monitoring, optimizing research, reducing rework and information fragmentation. From a cohesive database, specific and interrelated studies become possible, and this tool is designed to be free and open-source to help non-profit institutions that help people with Parkinson's disease.