Spatio-temporal distribution, prediction and relationship of three major acute cardiovascular events: Out-of-hospital cardiac arrest, ST-elevation myocardial infarction and stroke.

BACKGROUND

Predicting the incidence of time-sensitive cardiovascular diseases like out-of-hospital cardiac arrest (OHCA), ST-elevation myocardial infarction (STEMI), and stroke can reduce time to treatment and improve outcomes. This study analysed the spatio-temporal distribution of OHCAs, STEMIs, and strokes, their spatio-temporal correlation, and the performance of different prediction algorithms.

METHODS

Adults who experienced an OHCA, STEMI, or stroke in Canton Ticino, Switzerland from 2005 to 2022 were included. Datasets were divided into training and validation samples. To estimate and predict the yearly per-capita population incidences of OHCA, STEMI, and stroke, the integrated nested Laplace approximation (INLA), machine learning meta model (MLMM), the Naïve prediction method, and the exponential moving average were employed and compared. The relationship between OHCA, STEMI, and stroke was assessed by predicting the incidence of one condition, considering the lagged incidence of the other two as explanatory variables.

RESULTS

We included 3,906 OHCAs, 2,162 STEMIs, and 2,536 stroke patients. INLA and MLMM yearly predicted incidence OHCA, STEMI, and stroke at municipality level with very high accuracy, outperforming the Naïve forecasting and the exponential moving average. INLA exhibited errors of zero or one event in 82%, 87%, and 72% of municipalities for OHCA, STEMI, and stroke, respectively, whereas ML had errors in 81%, 89%, and 71% of municipalities for the same conditions. INLA had a prediction error of 0.87, 0.77, and 1.50 events per year per municipality for OHCA, STEMI and stroke, whereas MLMM of 0.70, 0.74, and 1.09 events, respectively. Including in the INLA model the lagged absolute values of the other conditions as covariates improved the prediction of OHCA and stroke but not STEMI. MLMM predictions were consistently the most accurate and did not benefit from the inclusion of the other conditions as covariates. All the three diseases showed a similar spatial pattern.

CONCLUSIONS

Prediction of incidence of OHCA, STEMI, and stroke is possible with very high accuracy using INLA and MLMM models. A robust spatio-temporal correlation between the 3 pathologies exists. Widespread implementation in clinical practice of prediction algorithms may allow to improve resource allocation, reduce treatment delays, and improve outcomes.