Establishment and validation of an artificial intelligence-based system for identifying the culprit vessel in patients with ST-segment elevated myocardial infarction: the ALERT study
We developed an AI algorithm to automatically identify the culprit vessel in patients with STEMI using 698 ECGs with confirmed culprit vessel. Through training, testing and validations, as well as real-world comparative testing, we demonstrated the effectiveness of this algorithm. Internal testing and external validation revealed that it achieved high performance metrics for identifying the culprit vessel. Comparison with experienced cardiologists also confirmed the competitive accuracy of the AI algorithm, especially for identifying LCX. Our study also provided preliminary evidence that the culprit vessel in patients with STEMI can be identified more accurately by designing appropriate neural network models with a small amount of data (online supplemental file 1).
In recent years, AI has been increasingly used in the field of medicine, especially in predicting and identifying CVDs.17–19 For qualitative diagnosis of STEMI, numerous ECG-based AI algorithms have been developed, such as the study conducted by Chang et al20 and our previous study4; however, there are fewer studies that advance to automatically identify the culprit vessel by analysing ECGs. In published studies, limitations such as improper model selection, lack of external validation or real-world comparative testing and insufficient outcomes of algorithms (especially for LCX) still exist.7–9 Compared with the study by Herman et al,21 which focused on detecting acute occlusion myocardial infarction, our algorithm focuses on identifying the culprit vessel in patients with STEMI, with a difference in clinical application scenarios. And compared with the study by Demolder et al,22 our study focuses on signal analysis rather than data conversion, but the fully automated high-precision ECG image processing technology proposed in that research offers the possibility of high-quality data input for our algorithm—particularly enabling real-time ECG analysis via smartphones in resource-limited areas. Besides, our Res-LSTM model offers advantages such as its efficacy with a small sample size and its suitability for medical image segmentation. Moreover, it can conserve computational costs and time without compromising predictive ability and demonstrate robust performance with time series data.9 23
In clinical practice, the prognosis of patients with STEMI is closely related to both the timeliness of treatment24 25 and accurate identification of the culprit vessel.26 Our previous study aids in addressing the issue of treatment delay.4 However, the identification of the culprit vessel, especially for patients with inferior wall infarction, remains clinically challenging.27 28 To address this problem, we developed the AI algorithm in this study to help identify the culprit vessel. This algorithm had the following advantages and potential applications for STEMI treatment. First, for individuals equipped with wearable ECG monitors and at high risk of AMI, the algorithm can be integrated into it to provide automatic qualitative and positional diagnosis, which may reduce the patient delay. Second, on patient’s arrival at the emergency centre, our AI algorithm can promptly analyse the ECGs, enabling clinicians to identify the culprit vessel and select the proper catheter in advance.29 Third, our algorithm may potentially shorten intraoperative reperfusion time. A recent study30 showed that patients with STEMI undergoing percutaneous coronary intervention (PCI) followed by CAG reduce operation time by approximately 6 min, with no significant differences in complications or prognosis. However, the primary challenge in performing PCI first is the lack of precise identification of the culprit vessel by clinicians. And our AI algorithm may aid in addressing this crucial issue. Fourth, accurate localisation of the culprit vessel can largely help the inexperienced clinicians in the Cardiac Critical Unit and interventional cardiologists, especially in the community hospital. Furthermore, AI algorithms can be continuously trained based on the growing database, which may potentially identify subtle ECG features to provide a more accurate diagnosis. In the future, the algorithm we developed may play a critical role in clinical decision-making in routine clinical practice.30 31
This study has certain limitations: First, most sampled patients experienced their first STEMI, with fewer cases of recurrent STEMI from previous MI, thus further research is needed for improving the accuracy. Second, the sampled patients had only one culprit vessel; therefore, further research is required to simultaneously identify patients with multiple culprit vessels. Third, the algorithm was only sampled and validated in China and has not been validated in other countries and races, so broader sampling and training is required for application on a global scale. Finally, this was a retrospective diagnostic study. More prospective and randomised controlled trials are warranted to further verify the clinical application, safety and effectiveness of this algorithm. Future work should address these limitations to make more progress in this field.
