Desenvolvimento de algoritmo de inteligência artificial para detecção de hemorragias intracranianas pós-traumas cranioencefálicos e seus potenciais benefícios no SUS Fácil.

Abstract

The study and use of Artificial Intelligence (AI) in medicine has grown exponentially. Auxiliary tools for identifying potential threats to life and prioritizing cases could be beneficial in the management of Traumatic Brain Injuries (TBIs) within the scope of the Unified Health System (SUS). These are public health problems, often resulting in Intracranial Hemorrhages (ICH) with high morbidity and mortality and socioeconomic costs, especially when not treated in a timely manner. In Minas Gerais, when they focus on cities with low hospital complexity, without neurosurgeons, radiologists, or computed tomographies (CTs), they need to transfer their patients, using the SUS FÁCIL bed regulation platform, through a hierarchical mechanism. The delay in the identification and registration of serious cases at the origin, their analysis by regulators in the SF regulation centers, could be alleviated with the potential insertion of an AI algorithm capable of automatically recognizing HICs in CTs at the origin sites, thus adding agility for the detection and prioritization of life-threatening cases. This work proposes the development of an ICH identification algorithm in CT and the analysis of its potential benefits in the context of SUS FÁCIL. After collecting and processing data sets images containing normal and HIC skull CTs, Orange® Software was used for training, validation and testing of their recognition by Artificial Neural Network (ANN) models, Support Vector Machines (SVM) and K – Nearest-Neighbors (KNN). The ANN model presented slightly better results than the other models in the Training / Validation and Testing stages (AUC = 1,000, CA = 0.998, F1 = 0.998, Accuracy = 0.998 and Recall = 0.998, and AUC = 0.987, CA = 0.930, F1 = 0.930, Precision = 0.931 and Recall = 0.930 respectively) achieving the proposed objective for the set of images used. However, the formulation of more robust datasets becomes necessary for their practical use and their improvement and use should be encouraged.