Screen Positives for Potential ST-Segment Elevation Myocardial Infarction: Door-to-ECG Time and Machine Learning

International guidelines require an ECG within 10 minutes for patients presenting with symptoms warranting  ST-segment elevation myocardial infarction (STEMI) screening. Delays in door to electrocardiogram (D2E) are associated with worse outcomes. Prior work identified D2E disparities associated with advancing age, Black race and female sex. Our objective was to determine the importance of social and behavioral determinants of health (SBDH) on D2E using a machine learning (ML) approach.

This retrospective study measured D2E for patients that screened positive based on standardized STEMI ECG screening criteria in a health system (including 11 EDs and > 500,000 ED visits per year). The cohort included walk-in patients that received at least one ECG in the ED from 1/1/2021-7/31/2023 and excluded encounters with D2E > 30 minutes or missing data. We utilized machine learning (XGBoost) to determine the predictive importance of age, sex, race, primary language, chief complaint (CC) of chest pain (CP), arrival period, as well as 17 domains of SBDH using an algorithm developed by our group. Predictive importance is reported as mean SHapley Additive exPlanations (SHAP) value, which provides the marginal contribution of a feature to the overall prediction of a D2E >10 minutes (positive values indicated increased likelihood and negative values indicate decreased likelihood). Variable interactions are reported as sumGain, where higher values indicate more influential interaction.

A total of 104,326 patient encounters met the inclusion criteria. Of these, 56.3% are female, 29% are Black with a median of 61 years of age. The most influential predictors and SHAP values of D2E > 10 min were CC of CP (-0.282), Black race (+0.112), age (+0.075), male sex (-0.054), arrival period business hours (+0.017), opioid abuse (0.012), nicotine dependence (+0.011), arrival period weekday night (+0.011). Interaction analysis revealed that age x Black race is the most influential (sumGain 316), followed by opioid abuse x Black race (sumGain 69) and age x nicotine dependence (sumGain 34).

With the addition of 17 SBDH domains, the CC of CP and non-modifiable patient factors (race, age, sex) remain the most predictive of D2E for those who screen positive for warranting an ECG. Future work will explore if SBDH factors may contribute to D2E disparities among those who screen positive and did not receive an ECG.