[ad_1]
Discussion
In this study, we developed and validated a CNN-based AI model for diagnosing NSTE-ACS in chest pain patients in the prehospital setting, using only the ECG as input. The AI model outperformed ECG interpretation by EMS paramedics, especially in identifying patients at low risk for having NSTE-ACS. Compared with a prehospital clinical risk score, the AI model did not match the diagnostic performance of the preHEART score. However, clinical risk scores could benefit from incorporating AI-enhanced ECG interpretation (figure 4).
Central image. ACS, acute coronary syndrome; AI, artificial intelligence; AUROC, area under the curve; ECG-AI, ECG interpretation by CNN; ECG-EMS, ECG interpretation by EMS paramedic; EMS, emergency medical services; NSTE-ACS, non-ST-elevation ACS; preHEART, prehospital History, ECG, Age, Risk factors and POC-troponin; STEMI, ST-segment elevation myocardial infarction.
To the best of our knowledge, this study stands out from prior research endeavours by our development and validation of a CNN-based AI model to classify NSTE-ACS based on the ECG. Moreover, we have undertaken a comparative analysis to assess the diagnostic performance of our model to the presently available prehospital diagnostic tools. Using AI-assisted interpretation of the ECG, the EMS paramedic is better able to differentiate between NSTE-ACS and non-cardiac chest pain compared with relying on their own interpretation. Especially in identifying patients at low risk for having NSTE-ACS, the AI model outperformed the EMS paramedic. Correctly identifying patients who are at low risk for having NSTE-ACS is crucial for optimal triage in the prehospital setting. Recent studies have shown that these low-risk patients could potentially be left at home or transferred to a general practitioner, leading to less ED overcrowding and lower healthcare costs.11 Nevertheless, the PPV of the AI model is only modest and does not outperform ECG interpretation by EMS paramedics. This is likely due to the heterogeneity within the NSTE-ACS population, as some patients present with significant ischaemic changes on the ECG, while others show minimal or no changes. This variability makes consistent risk stratification based on the ECG findings challenging. In addition to improving diagnostic performance, using a model for ECG interpretation in the prehospital setting addresses the issue of inter-operator variability. Inter-operator variability is common among medical personnel when scoring the ECG and is highly dependent on experience and training.6 7 31 EMS paramedics receive training aimed at recognising a STEMI, but no training in further assessing more subtle ECG changes such as minimal ST-depression and T-wave inversion.32 33 However, these changes are important to recognise and score in a clinical risk score.5 15
Other studies, which used a machine learning-based AI model for prehospital ECG diagnosis in patients with chest pain, demonstrated better overall diagnostic performance compared with our AI model.22 23 However, these studies were aimed at a study population with a high prevalence of occlusive myocardial infarction (OMI), which could explain the differing levels of diagnostic performance.22 In patients with OMI, ischaemia is persistent, making ECG abnormalities more pronounced and therefore more easily recognised by medical experts and AI models alike.22 34 Moreover, in our study population, we included a more selected population as patients with suspected NSTE-ACS with very high risk criteria according to the ESC and AHA/ACC guidelines (eg, haemodynamic unstable, cardiogenic shock, ongoing chest pain refractory to medical treatment) in the EMS setting were excluded.1 2 Therefore, by excluding very high-risk patients on one hand and including a substantial number of ACS patients with non-occluded myocardial infarction on the other, we likely encounter more subtle, if any, ECG abnormalities, which also limit detection by an AI model.
Previous studies have shown that clinical risk scores have the best diagnostic performance for prehospital risk stratification in suspected NSTE-ACS patients, and our results confirm these findings.5 8 Our study shows that using single prehospital diagnostic tools, such as ECG and POC-troponin, both by human interpretation and AI, is not sufficient when used alone in the prehospital setting to identify NSTE-ACS patients.35 However, combining clinical parameters in a clinical risk score significantly improves the overall diagnostic performance. Additionally, incorporating AI into clinical risk scores can enhance diagnostic performance even further.
Limitations
This study suffers from limitations inherent to the study design. First, we used only DeepArrNet to develop the CNN-based model. We did not investigate alternative architectures to determine whether they might yield a model with superior diagnostic performance. Second, the AI model was developed within a study population of modest size. Nonetheless, despite this limited cohort, we observed a noteworthy enhancement in ECG interpretation accuracy in patients at low risk for having NSTE-ACS. It is anticipated that further refinement can be achieved through the training of the model in larger populations, thereby augmenting its overall performance. A third limitation is that the ECG-AI score is only validated in a very specific cohort of patients within the EMS setting. The model’s reliability in patients without chest pain or in other settings remains unknown. Although the AI model demonstrated superior diagnostic performance compared with ECG interpretation by EMS paramedics, its effectiveness should be evaluated in other settings and healthcare systems with potentially different chest pain populations to determine whether this performance advantage persists.
[ad_2]
Source link
