Co-recorded conditions in ischemic and hemorrhagic stroke: a 20-year national analysis of 598,186 natural deaths with machine learning.

The purpose of this study was to characterize the co-occurrence patterns of co-recorded conditions in ischemic and hemorrhagic stroke and identify co-recorded conditions using a large national dataset of fatal, subtype-confirmed stroke cases from Brazil. We analyzed 598,186 natural deaths recorded in the Brazilian Mortality Information System (SIM) between 2000 and 2019, comprising 288,724 (48.3%) hemorrhagic and 309,462 (51.7%) ischemic stroke. Hierarchical clustering was performed to identify co-recorded condition clusters, while association rule mining using the Apriori algorithm uncovered hidden patterns of disease co-occurrence. Logistic regression, random forest, and XGBoost models were developed to classify stroke subtype based on demographic characteristics and co-recorded condition profiles. Ischemic stroke patients had significantly higher co-recorded condition burden (median 2, IQR 1-3) compared to hemorrhagic stroke (median 1, IQR 1-2; p < 0.001). For hemorrhagic stroke, 12 association rules were identified; the strongest involved co-occurrence of infectious and respiratory diseases (e.g., {hypertension = Yes, infectious_disease = Yes} → {respiratory_disease = Yes}, lift = 2.57) and of endocrine/metabolic conditions with hypertension (lift = 1.67), whereas the rule {mental_disorders = Yes} → {hypertension = Yes} (lift = 1.27, confidence = 53.9%) was the weakest of the 12 rules. In contrast, ischemic stroke yielded 8 rules reflecting associations among infectious, respiratory, genitourinary, and endocrine/metabolic conditions. The logistic regression model achieved an AUC of 0.768 for predicting stroke subtype, with consistent performance across sensitivity analyses. Age-standardized analyses showed that ischemic stroke proportion increased from 48.8 to 51.5% over 20 years, with approximately half of the crude increase attributable to population aging (crude ischemic proportion reached 54.7%). Co-recorded condition burden increased substantially in both subtypes (≥3 co-recorded conditions in ischemic stroke: 24.8-33.3%). We observed that hemorrhagic and ischemic stroke have different co-occurrence patterns of co-recorded conditions: hemorrhagic stroke exhibited simpler co-occurrence patterns centered on hypertension together with infectious, respiratory, and endocrine/metabolic conditions, while ischemic stroke was characterized by clustered cardiovascular and metabolic multimorbidity. These findings warrant further investigation in prospective studies.