From detection to deployment: AI for culprit vessel identification and the path to practice

Ischaemic heart diseases (IHDs) encompass a range of conditions, among which ST-segment elevation myocardial infarction (STEMI) represents a particularly severe form. STEMI is characterised by a complete occlusion of a coronary artery, and it benefits from rapid revascularisation through primary percutaneous coronary intervention (PCI) within 120 min from ECG-based diagnosis.1 Meeting these time requirements not only necessitates a rapid diagnosis but also quick identification of the culprit vessel so that appropriate treatment can be delivered promptly.

Rapid diagnosis and thereby intervention times depend on several factors, notably the recognition of clinical symptoms, such as chest pain, and the identification of specific patterns on a 12-lead ECG. Nevertheless, the correct interpretation of the ECG and the identification of the culprit vessel are influenced by multiple factors, including variations in coronary artery anatomy,2 examination conditions and, crucially, the physician’s experience. It has been suggested that differences in ECG interpretation may exist between junior and senior physicians, as well as between cardiologists and other specialists.3

The employment of artificial intelligence (AI) to interpret ECG waveforms and facilitate the identification of the responsible artery may provide a useful tool to diminish variation in time to diagnosis and, consequently, reduce delays in initiating treatment. Research has found that machine learning (ML) applications for ECG-based STEMI diagnosis often lack information on the culprit vessel.4 Additionally, models often struggle with detecting blockages in the left circumflex (LCx) artery,5 which remains the most difficult to identify.

A study by He et al. 3 established and validated an AI-based algorithm capable of identifying the culprit vessel through 12-lead ECG analysis. Their dataset comprehensively included information on various blood vessel positions, also incorporating cases of patients with LCx artery blockages, which are often under-represented in other studies. Their approach identifies the culprit vessel across all the different positions. During the validation phase, cardiologists outperformed the model in diagnosing left anterior descending artery occlusions. However, the model showed particular strength in identifying occlusions in the LCx and right coronary artery.

Notably, LCx occlusion cases represent only 20% of the dataset. Despite the limited representation, the results are promising, particularly considering that the diagnosis of this artery through ECG is notoriously difficult for clinicians. Indeed, approximately one-third of patients with acute coronary syndrome (ACS) with LCx occlusion exhibit no ST-segment changes in ECGs.6

The limitations of the article by He et al. 3 are mainly related to the characteristics of the population studied. The study was conducted using data from hospitals located exclusively in China, which limits the generalisability of the results to other regions. The dataset includes patients with STEMI only. Since the model’s primary goal is to identify culprit vessels, expanding it to encompass a broader spectrum of ACS would be beneficial, moving beyond the current differentiation between STEMI and NSTEMI (non-STEMI) cases. This will allow for the inclusion of more diverse clinical scenarios, where ECG changes are less distinct and specific than in STEMI, potentially improving the model’s diagnostic capabilities.

The application of AI methods to enhance ECG analysis, particularly for culprit vessel identification, can lead to faster and more efficient diagnosis by potentially reducing misinterpretation and minimising time delays,7 ultimately resulting in faster interventions when necessary. However, realising these benefits in routine clinical practice requires more than promising results, as clinical decision-making in medicine is typically grounded in a hierarchy of evidence.8

As described in a study by Bellini et al.9, successful clinical integration must be supported by best practices, beginning with internal validation, followed by external one, and concluding in outcome evaluation. In this context, further validation is essential, especially external one across different populations and clinical settings. For instance, assessing performance in emergency departments with diverse ECG equipment or different patient demographics would help improve the algorithm generalisability. Randomised controlled trials also play an important role. They help determine whether identification of the culprit vessel via AI can improve clinical outcomes, for example, by reducing time to reperfusion in STEMI cases. Transparency through open-source code and reproducible methods is equally important to build trust and support widespread adoption of AI tools10 in cardiovascular care.

To conclude, bridging the gap between research and real-world implementation is not just about creating better models. For AI tools to effectively improve care, we need to focus not only on their design and validation but also on the path towards trustworthy implementation. This becomes particularly important in situations that need prompt and precise decision-making, such as identifying the culprit vessel for IHDs.

Data availability statement

No data are available.

Ethics statements

Patient consent for publication

Not applicable.

Ethics approval

Not applicable.