Protocolo de classificação com inteligência artificial de pacientes com quadro de AVC agudo no contexto da emergência

Abstract

Context: Among the health problems that affect the population, cerebrovascular accident (CVA) is one of the most serious, as it is a disease with significant mortality and morbidity. Prompt recognition and treatment of this disease are necessary to reduce the impacts on the affected individual's health. This requires imaging exams and immediate interpretation by professionals trained for this purpose, who are not always present in health units. With the advent of Artificial Intelligence (AI) technology, new computational support tools for medical diagnosis (computer aidede diagnosis – CAD) emerged, enabling agility and improved performance in this task. Objective: This study aims to present and validate a tool for classification of patients with cerebrovascular accident conditions using a solution based on Artificial Intelligence called SDAC-AVC. This tool consists of two classifiers: one based on clinical findings (semantic) and another based on image characteristics (Convolutional Neural Network-RNC). Method: 100 cases of patients affected by cerebrovascular accident were evaluated, 67 cases of the ischemic type and 33 cases of the hemorrhagic type, all submitted to clinical evaluation and CT scan of the skull. The exams were evaluated by the specialist and submitted to classification by the AI tool according to the type of injury present, whether ischemic or hemorrhagic stroke, and then the quality metrics were evaluated: Sensitivity, Specificity, Accuracy, Positive Predictive Value (PPV) and Predictive Value Negative (VPN). Results: the results found for the semantic classifier in the classification of ischemic lesions were sensitivity 0.58, specificity 0.54, accuracy 0.57, PPV 0.12 and NPV 0.92. For lesions of a hemorrhagic nature, the values found were: sensitivity 0.54, specificity 0.58, accuracy 0.57, PPV 0.12 and NPV 0.92. For the RNC classifier, the values found for ischemic lesions were sensitivity 0.91, specificity 0.96, accuracy 0.93, PPV 0.28 and NPV 0.99. For hemorrhagic lesions, the values found were sensitivity 0.96, specificity 0.91, accuracy 0.93, PPV 0.12 and NPV 0.99. The results were compared to those of other tools already existing and available on the market, achieving similar performance. Conclusion: The AI tool based on clinical findings showed an unsatisfactory performance in the classification of stroke cases, and its use is not advisable. On the other hand, the NCR-based tool for classification of findings from imaging exams in patients affected by stroke showed a very satisfactory performance, and its application in the context of Emergency is an option for diagnostic aid to the on-duty physician.