Human–Computer Agreement of Electrocardiogram Interpretation for Patients Referred to and Declined for Primary Percutaneous Coronary Intervention: Retrospective Data Analysis Study (Preprint)

Abstract

BACKGROUND When a patient is suspected of having an acute myocardial infarction (AMI), patients are accepted or ‘turned-down’ for primary percutaneous coronary intervention (PPCI) partly based on clinical assessment of the 12-lead electrocardiogram (ECG) and STEMI criteria. OBJECTIVE This paper aims to retrospectively interrogate the agreement rate between the human (specialist nurse called PPCI ‘activator’) and computer interpretations of ECGs for cases where patients are turned-down for PPCI. METHODS Various features of patients who were referred for PPCI were analysed. Both the human and computer ECG interpretations were simplified to either being suggestive or not suggestive of AMI to avoid analysis of complex heterogeneous and synonymous diagnostic terms. Analysis to measure agreement was carried out along with logistic regression to measure the association between these interpretations (and other variables) in predicting mortality. RESULTS A total of 1464 patients referred for PPCI were turned-down. For these cases, 722 (49%) of the computerised diagnoses were suggestive of AMI whereas 634 (43%) of the human interpretations were suggestive of AMI (p<0.001). The human and computer agreed that there was a possible AMI in 342 cases (23.3%). However, both agreed more often that the patient was not having an AMI (n=450 [31%]). The overall agreement rate was 54%. Cohen's kappa coefficient (κ) showed a poor agreement (κ= 0.08 [p=0.001]). Other factors including age (odds ratio=1.06, 1.07) and chest pain (odds ratio=0.47, 0.59) were the only independent variables that were statistically significant (p<0.01) in predicting mortality after 30 days and 1 year. The odds for mortality within 1 year of referral were lower in patients with chest pain compared to those patients without chest pain. A referral being ‘out of hours’ is a ‘trending’ variable (odds ratio=1.93, 1.41, p=0.09) for predicting the odds of 1-year mortality. Other insights related to human and computer performance over hours of the day are also provided. CONCLUSIONS Mortality in patients turned down for PPCI is higher than the reported mortality for STEMI patients at one year. Agreement between computerised and human ECG interpretation is poor perhaps leading to a high rate of inappropriate referrals. Work is needed to improve machine and human decision making when reading ECGs to ensure that patients are signposted to the correct treatment facility for time-critical therapy.