Dados do Trabalho

Título

VALIDATION OF AN ARTIFICIAL INTELLIGENCE IN THE PREDICTION OF STROKE RECURRENCE: PRELIMINARY RESULTS

Resumo

Introduction: Stroke remains the main cause of death in Ceará (Brazil). The high number of deaths in the state reflects the large incidence of primary and recurrent strokes. With the advances in digital health, technologies such as Artificial Intelligence (AI) can help the development of prediction studies. Machine Learning, an area of AI, is defined as the ability of a machine to learn “by itself” through database analysis and perform predictions, such as the recurrence of a stroke in a timely manner with minimal error. Objective: To validate an artificial intelligence that helps predict the recurrence of stroke. Method: The database used for training the predictive algorithm consolidated the records of 571 individuals residing in Ceará. It was built from three databases, with 127 records acquired with the Neurofunctional Physiotherapy League of the Federal University of Ceará database; 66 records by a survey from the Fortaleza General Hospital; and 378 records, from a population database, by an internet form survey. The variables included in the model were: occurrence of stroke, sex, age, marital status, hypertension, heart disease, type of work, local of residence, diabetes mellitus, tobacco use, alcohol use, and depression/anxiety. Results: These are preliminary results of an ongoing study. Of the 571, 237 people had a primary stroke and 51 of those had a recurrence. The algorithm with the best predictive performance was the Light Boosting Gradient Model (LBGM) with the highest accuracy of 90.0% and an AUC (Area Under the Curve) of 96.5%. The higher the accuracy and the AUC, the more positive model’s accuracy. Among the variables, those with high influence scores on the stroke recurrence were: 1st stroke (80.1), heart disease (76.2), age > 50 years (48.5), Hypertension (46.9), BMI > 24 (18.4), Diabetes mellitus (12.5), Alcohol use (8.5), Depression/Anxiety (5.18), urban residents (4.77), married (3.03), retired (2.81), male (0.61) and smoking (0.26). Individuals after the first stroke frequently stopped smoking, this shows that the change in this habit makes it less explicative in the model. Conclusion: The AI algorithm is valid and has accuracy in stroke recurrence prediction models, with LBGM, a modern machine learning algorithm, being the most accurate. According to the analysis, the first stroke makes the individual susceptible to a second stroke, and the change in modifiable factors may be related to a lower influence of recurrence of stroke.

Palavras Chave

Stroke; Artificial Intelligence; Recurrence; Forecasting; Machine Learning