Prediction of emergency department presentations for acute coronary syndrome using a machine learning approach

Abstract

The relationship between weather and acute coronary syndrome (ACS) incidence has been the subject of considerable research, with varying conclusions. Harnessing machine learning techniques, our study explores the relationship between meteorological factors and ACS presentations in the emergency department (ED), offering insights into seasonal variations and inter-day fluctuations to optimize patient care and resource allocation. A retrospective cohort analysis was conducted, encompassing ACS presentations to Dutch EDs from 2010 to 2017. Temporal patterns were analyzed using heat-maps and time series plots. Multivariable linear regression (MLR) and Random Forest (RF) regression models were employed to forecast daily ACS presentations with prediction horizons of one, three, seven, and thirty days. Model performance was assessed using the coefficient of determination (R²), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The study included 214,953 ACS presentations, predominantly unstable angina (UA) (94,272; 44%), non-ST-elevated myocardial infarction (NSTEMI) (78,963; 37%), and ST-elevated myocardial infarction (STEMI) (41,718; 19%). A decline in daily ACS admissions over time was observed, with notable inter-day (estimated median difference: 41 (95%CI = 37-43, p = < 0.001) and seasonal variations (estimated median difference: 9 (95%CI 6-12, p = < 0.001). Both MLR and RF models demonstrated similar predictive capabilities, with MLR slightly outperforming RF. The models showed moderate explanatory power for ACS incidence (adjusted R² = 0.66; MAE (MAPE): 7.8 (11%)), with varying performance across subdiagnoses. Prediction of UA incidence resulted in the best-explained variability (adjusted R² = 0.80; MAE (MAPE): 5.3 (19.1%)), followed by NSTEMI and STEMI diagnoses. All models maintained consistent performance over extended prediction horizons. Our findings indicate that ACS presentation exhibits distinctive seasonal changes and inter-day differences, with marked reductions in incidence during the summer months and a distinct peak prevalence on Mondays. The predictive performance of our model was moderate. Nonetheless, we obtained good explanatory power for UA presentations. Our model emerges as a potentially valuable supplementary tool to enhance ED resource allocation or future predictive models predicting ACS incidence in the ED.

Figure 1

Temporal pattern of ACS emergency department visits. The temporal trends and fluctuations were visualized by plotting 30-day moving averages of daily emergency department visits for ACS and its subdiagnoses. ACS Acute coronary syndrome, ED emergency department, UA unstable angina pectoris, NSTEMI non ST-elevated myocardial infarction, STEMI ST-elevated myocardial infarction.

Figure 2

(a) Heat-map visualizing the sum of absolute ED presentations per weekday. (b) Fig. 2b Heat-map visualizing the sum of absolute ED presentations per month.

Figure 3

(a) Feature importance of Random Forest model. (b) Feature importance of the LASSO model. The top 10 importance of individual predictors for machine learning algorithms predicting ACS incidence with a one-day prediction horizon. Abbreviations: LASSO: Least Absolute Shrinkage and Selection Operator. LASSO was employed concurrently with the Multivariable Linear Regression model to serve as a feature selection tool.