Artificial intelligence in ECG analysis: current practice and future directions from the international SAIMO physician survey

Background While many artificial intelligence (AI) algorithms are already available for the analysis and interpretation of electrocardiogram (ECG) signals, it is unclear if these tools have earned the trust of healthcare personnel. Purpose To assess the use and level of trust in AI algorithms for ECG analysis in routine clinical practice. Methods This is the preliminary analysis of the "Survey on Artificial Intelligence in heart rhythm Monitoring" project, an online survey that assesses healthcare professional’s attitude toward the use of AI in ECG signal interpretation. Respondents’ trust in AI algorithms was tested in three cases: Case 1 described a young patient with palpitations suggestive of supraventricular tachycardia during 14-day ECG recording and a negative result from AI-based ECG signal analysis. Case 2 describes a patient with coronary artery disease, previous syncope who experienced symptoms during 14-day ECG recording, again with a negative AI result. Case 3 described a patient with acute coronary syndrome, without criteria for ST-segment elevation myocardial infarction but with an AI alert suggesting obstructive myocardial infarction. Results A total of 203 experts (46.3% cardiologists, 42.3% cardiac electrophysiologists, 11.3% other specialists; 24% women; median age 40-49 years) from 33 countries and 7 continents participated in the survey to date. The majority of respondents stated that a prospective randomized trial (46.5%) or a prospective multi-center cohort study (37.0%) should be conducted before approving an AI algorithm for clinical use. Respondents found AI assistance appropriate in the analysis of 14-day ECG (73.0%), wearable-detected ECG (71.5%) and 12-lead ECG (73.5%). AI-only analysis without the possibility to manually review the ECG signal was deemed suitable in only the minority of cases. The most suitable scenarios for AI-only analysis were the detection of atrial fibrillation (26.5%) and the monitoring of a young, healthy patient with palpitations (24.0%). In case of an AI algorithm missing a critical arrhythmia, respondents recommended that both the physician using the algorithm (44.0%), the manufacturer of the algorithm (36.5%) or the healthcare provider (15.0%) should be held accountable. In Case 1, 95.5% of experts would not trust the AI algorithm and either analyze the ECG signal manually (60.8%) or perform further rhythm monitoring (34.7%). In case 2, 97.5% would not trust the ECG analysis, they would analyze the ECG signal again manually (56.7%) or repeat rhythm monitoring (40.7%). In case 3, 12.1% of experts followed the AI algorithm’s suggestion of immediate cath lab activation. The majority (79.3%) would recommend transfer to the emergency department for individual evaluation of the patient. Discussion: This survey shows that experts in ECG analysis have only limited trust in currently available AI algorithms, especially in case of results that diverge from clinical suspicion.