The Use of Artificial Intelligence-Enhanced Electrocardiograms in the Chest Pain Clinic to Risk Stratify Patients, Provide Rapid Reassurance and Enable a Low-Cost Clinical Pathway

Our group has been at the forefront of developing AI-ECG models that are actionable, biologically plausible and explainable, with the aim of fostering clinical trust and maximising their potential for integration into routine clinical care. Recently, we developed the artificial-intelligence risk-estimation (AIRE) platform, comprising a series of AI-ECG models capable of predicting time-to-mortality, and further disease-specific AIRE models encompassing a broad range of future cardiovascular conditions from a single 12-lead ECG.

The AIRE platform was developed using the Beth Israel Deaconess Medical Centre (BIDMC) dataset, comprising more than 1.1 million ECGs from a secondary care population in Boston, USA. The tool has been externally validated across five large international cohorts from the USA, UK, and Brazil, encompassing diverse demographics, a range of baseline cardiovascular risks, including both primary and secondary care, as well as volunteer populations.

AIRE generates patient-specific survival curves using data from a single 12-lead ECG and can predict time-to-death. In our validation studies, AIRE demonstrated strong predictive performance for cardiovascular (CV) mortality, achieving a concordance-index (c-index) of 0.844 (95% confidence interval, 0.839-0.849). This outperformed current predictive models based on demographic data and traditional risk factors alone, which had a c-index of 0.733 (0.726-0.740).

We believe the rapid access chest pain clinic provides an ideal setting in which to evaluate the clinical translation of this validated AI-ECG technology. Risk prediction is central to clinical-decision making in this population, making it well-suited for assessing the added value of AI-enhanced tools. In low-risk cohorts such as our proposed study population, the high negative predictive value of AIRE offers the potential to confidently reassure patients at the lowest risk, in line with national guidelines. Importantly, we have demonstrated that AIRE can stratify risk even among patients whose ECGs are labelled as normal by clinicians. In this subgroup, the model effectively distinguishes between low- and higher-risk individuals, with a significant difference in mortality observed between these groups.

ORIGINAL HYPOTHESIS

Primary hypothesis

1. An AI-ECG-assisted chest pain pathway will result in lower healthcare resource cost compared to standard clinical investigation pathways.

   Secondary hypotheses

2. An AI-ECG-assisted chest pain pathway in patients with non-cardiac sounding chest pain is non-inferior to routine care in terms of:

   1. preventing cardiovascular events and mortality
   2. initiating standard primary prevention medications
   3. achieving overall symptom resolution
3. An AI-ECG-assisted chest pain pathway in patients with non-cardiac sounding chest pain is superior to routine care in reducing the time from referral to diagnosis with a definitive management plan.

EXPERIMENTAL DETAILS AND DESIGN OF PROPOSED INVESTIGATION Proposal

1. There is a substantial and growing burden of referrals to chest pain clinics, often resulting in extensive investigations.
2. For patients with non-cardiac sounding chest pain, a referral and investigation pathway is unlikely to provide a significant clinical benefit over guideline-based care.
3. Novel AI-guided risk prediction tools have the potential to provide rapid, low-cost reassurance for patients unlikely to benefit from further investigation.

Aims To determine whether AI-ECG screening can facilitate reassurance of patients presenting to rapid access chest pain clinic with non-cardiac sounding chest pain, with less resource utilisation, without adversely affecting cardiovascular outcomes.

Study design This study is a multicentre, prospective, cluster-randomised trial comparing AI-ECG-assisted management with routine care in patients presenting to rapid access chest pain clinics with non-anginal chest pain.

Methodology The majority of rapid access chest pain clinics in North West London are led by specialist chest pain nurses, with support from Consultant Cardiologists. At the start of each week, three of the six participating clinic sites will be randomised to deliver either standard clinical care or an AI-ECG guided pathway. The remaining three sites will be assigned to the alternative pathway, ensuring that each week three sites will use each pathway. All patients meeting the inclusion criteria who present to the clinics during the study will follow the care pathway to which their clinic has been assigned.

In clinics following the AI-ECG guided pathway, patients will be assessed as usual by a clinician, and in addition, undertake a quantitative symptom questionnaire based on the Rose angina classification, which we are planning to embed into an app. This will identify the nature of chest pain symptoms and establish if the chest pain is in keeping with angina or not, in a reproducible, standardised and objective format. As previously discussed, atypical angina is a vague and poorly defined descriptor. Chest pain will be described as either being typical for angina or non-anginal. Patients with typical anginal pain will undergo standard clinical investigation as advised by the treating clinician.

All patients will undergo a 12-lead ECG that is analysed using our AI-ECG model. Patients with typical angina will have clinician determined further investigation without AI analysis feedback, but the AI-ECG predictions will be collated for later sub-analyses. For patients with non-anginal chest pain, the AI-ECG algorithm will risk stratify ECGs and identify a low-risk group. We have defined low risk as having a 2% or lower predicted risk of a major adverse cardiovascular events within 1-year according to the AI-ECG analysis.

Based on analysis of a comparable, age and sex matched cohort in our BIDMC dataset, we predict that around 60% of patients will fall into this low-risk group. Patients with non-anginal chest pain identified as having low-risk ECGs by our AI algorithm will be reassured with no further diagnostic testing.

All patients will undergo standard cardiovascular risk factor assessment and will be counselled on lifestyle modification. Where clinically indicated, primary prevention therapy, including statins, antihypertensives, and diabetic medications, will be recommended in accordance with national guidelines. In the SCOT-HEART study, which addressed the use of CT coronary angiography in a similar patient group, a key weakness was a lower implementation of primary prevention therapy in the control arm compared to the intervention group. We plan to adopt an aggressive approach to primary prevention therapies which we will maintain consistently across study groups, and measure adherence to as a secondary outcome of the trial.

At the end of each week, all clinical assessments and patient management pathways will be reviewed by a Consultant Cardiologist to ensure safety and adherence to clinical standards. The North West London Whole Systems Integrated Care (WSIC) digital information system enables linkage to primary and secondary care records, including presentation, diagnostic and treatment information, as well as mortality outcomes for the 2.3 million population of North West London. We will use WSIC data to assess healthcare utilisation, the occurrence of cardiovascular events, and the prescription of statins and antihypertensives in the 12 months following each patient's assessment.

A subgroup of 500 patients who have consented for follow up will be reviewed via telephone at one year to assess symptom status using the Rose Angina quantitative questionnaire app, consistent with the baseline assessment, and we will perform analysis of these patient-reported outcomes. In addition, we will evaluate patient experience and satisfaction with the assigned care pathway to gain insights into its acceptability, perceived effectiveness, and overall quality from the patient perspective. We propose the utilisation of the app will provide a standardised and objective format for structured analysis of patient reported symptoms, whilst also providing an open platform for qualitative data collection and overall thematic analysis.