Impact of automatic acquisition of key clinical information on the accuracy of electrocardiogram interpretation: a cross-sectional study

Background The accuracy of electrocardiogram (ECG) interpretation by doctors are affected by the available clinical information. However, having a complete set of clinical details before making a diagnosis is very difficult in the clinical setting especially in the early stages of the admission process. Therefore, we developed an artificial intelligence-assisted ECG diagnostic system (AI-ECG) using natural language processing to provide screened key clinical information during ECG interpretation. Methods Doctors with varying levels of training were asked to make diagnoses from 50 ECGs using a common ECG diagnosis system that does not contain clinical information. After a two-week-blanking period, the same set of ECGs was reinterpreted by the same doctors with AI-ECG containing clinical information. Two cardiologists independently provided diagnostic criteria for 50 ECGs, and discrepancies were resolved by consensus or, if necessary, by a third cardiologist. The accuracy of ECG interpretation was assessed, with each response scored as correct/partially correct = 1 or incorrect = 0. Results The mean accuracy of ECG interpretation was 30.2% and 36.2% with the common ECG system and AI-ECG system, respectively. Compared to the unaided ECG system, the accuracy of interpretation was significantly improved with the AI-ECG system (P for paired t-test = 0.002). For senior doctors, no improvement was found in ECG interpretation accuracy, while an AI-ECG system was associated with 27% higher mean scores (24.3 ± 9.4% vs. 30.9 ± 10.6%, P = 0.005) for junior doctors. Conclusion Intelligently screened key clinical information could improve the accuracy of ECG interpretation by doctors, especially for junior doctors.