Abstract
Background: Brazil has one of the highest stroke rates in Latin America. It is important to understand the impact of other causes of death and sociodemographic factors, as this may contribute to a better comprehension of the stroke mortality process. Machine learning provides a means to explain this process.
Objective: To investigate the stroke deaths profile and its subtype in Brazil using machine learning.
Methods: This is a time series analysis where deaths mentioning stroke and other conditions were identified using individual death records from the country’s mortality information system (SIM) between 2000 and 2019. Strokes were grouped into the following subtypes: ischemic stroke (IS), hemorrhagic stroke (HS), and unspecified stroke (US). A decision tree model was built to identify the strongest factors distinguishing IS from HS.
Results: There were 2,459,742 deaths mentioning stroke. There was a progressive increase in the number of deaths mentioning stroke over the study period. The most common type of stroke was US, accounting for more than 62% of deaths. Among HS deaths, hypertensive diseases were the most frequent group of associated causes (40.6%), while the most frequent group in subtypes IS and US was diseases of the respiratory system (48.30% and 42.30%, respectively). The decision tree analysis revealed that IS was more likely to occur in patients aged 60 years and over and in cases where respiratory diseases, endocrine diseases, arrhythmias, ischemic heart disease and heart failure were present. However, HS was more frequent in younger patients without these conditions but with nervous system diseases.
Conclusions: The decision tree analysis identified the strongest factors distinguishing IS from HS, highlighting variables involved in each subtype of stroke-related death that can be recognized in clinical practice. These variables may also support the redistribution of deaths initially classified as unspecified stroke.