Using AI to Diagnose MI

I’m writing this blog entry after submitting a narrative review for a journal client. I thought it might be fun to break down my findings and present a quick guide to the topic of applying AI to emergency chest pain patients. This blog is not fully sourced. Two key references are linked below as they form the basis of the arguments presented.

Background

Chest pain presents a particular challenge for clinicians in the emergency department. The majority of chest pain is caused by non-life-threatening conditions, such as musculoskeletal pain. However, failing to pick up and appropriately treat serious causes, like myocardial infarction (MI), can lead to potentially fatal consequences. A particular challenge is differentiating between Type 1 MI (caused by a blockage in the arteries supplying the heart) and Type 2 MI (due to other issues in heart muscle oxygen supply and demand).

The use of artificial intelligence (AI), and specifically machine learning (ML), has long been proposed as a potential solution to this diagnostic challenge. There is an extensive history of research into the use of AI in medicine, showing increasing promise with its ability to support accurate diagnosis and prioritise smart resource allocation. The application of such tools to the emergency chest pain patient has also been the subject of much research.

2024 saw the completion of the RAPIDxAI trial (1). This study represents a significant step forward towards the real-world application of such technology. Researchers employed the use of an AI tool to take various data points from patients suffering from chest pain in the emergency department, producing a risk of Type 1 MI assessment alongside a recommended treatment plan. Relevant health outcomes were then compared to normal care (without the use of AI). Disappointingly, the use of this tool did not lead to a reduction in negative health outcomes (such as readmission or a repeat MI). However, it is important to note that as outcomes were not any worse, this trial did provide excellent evidence for the safety of implementing AI tools in the clinical environment. Researchers also noted a beneficial effect on resource allocation, with patients undergoing fewer unnecessary invasive investigations when compared to normal care.

To fully understand the impact of the RAPIDxAI trial, it is important to place its findings in the context of prior research on the subject. One key paper in this field is Stewart et al.’s 2020 systematic review: Applications of machine learning to undifferentiated chest pain in the emergency department: A systematic review (2). Drawing conclusions from this paper and other key research since its publication, we are able to discuss some of the key strengths and limitations of current research and ultimately identify key next steps in advancing the use of these tools towards real-world clinical application.

Strengths of current research

Various ML models have consistently demonstrated a high level of sensitivity for diagnosing Type 1 MI (i.e. they very rarely fail to pick up when a Type 1 MI is happening). We have also gathered much high-quality evidence on which data points are most useful in improving ML model performance. Time and time again, AI has been able to provide accurate risk information to clinicians and prove its usefulness as a safe ‘rule-out MI’ tool. More recently, high-quality trials such as RAPIDxAI have focussed more on real-world outcomes and assessing their usefulness on the frontline. This also includes introducing AI to other parts of the care system besides diagnosis (such as supporting imaging or lab value interpretation).

Limitations of current research

However, research into the field still faces several significant limitations. Stewart et al. identified the lack of models trained on sufficiently large data sets, a process vital for improving accuracy and reliability. There is also limited research into the human factors behind clinicians' reluctance to adopt AI tools into clinical practice, something the RAPIDxAI authors noted in their analysis of the trial results. Progress in the field is consistently held back by the lack of readily available ML model protocols and algorithm training information. This lack of transparency makes reproducing research results highly challenging. Finally, recent studies have questioned the diagnostic specificity of AI applied to emergency chest pain patients (i.e. the ability of these tools to correctly identify people not having a Type 1 MI).

Next steps

Assessing recent findings alongside the broader context of prior research, we are able to identify the following important next steps research must take to move towards real-world application:

• Training of ML models on sufficiently large data sets.

• Addressing human factors barriers to clinician acceptance.

• Investigating long-term patient outcomes, particularly incorporating data from invasive procedure risk of harm.

• Undertaking a high-level ‘meta-analysis’ of current evidence to guide future research aims.

Once the full review has been published, I will link to the open-access paper here. Further details and source material will be available on publication or request.

Key Sources

1. Saha A. Re-engineering the Clinical Approach to Suspected Cardiac Chest Pain Assessment in the Emergency Department by Expediting Evidence to Practice Using Artificial Intelligence - RAPIDxAI. [online] American College of Cardiology Maintenance; 2024 [Accessed 2024 Nov 24]. Available from: https://www.acc.org/Latest-in-Cardiology/Clinical-Trials/2024/08/29/20/29/rapidxai

2. Stewart J, Lu J, Goudie A, Bennamoun M, Sprivulis P, Sanfillipo F, et al. Applications of machine learning to undifferentiated chest pain in the emergency department: A systematic review. Bivona G, editor. PLOS ONE. 2021 Aug 24;16(8):e0252612.