Development and validation of an AI decision-support tool to improve diagnosis and risk stratification of patients presenting in the Emergency Department

Background

Cardiovascular disease frequently presents with symptoms such as acute chest pain or breathlessness Together, individuals with these symptoms account for 1 in 4 presentations to the Emergency Department and 1 in 10 unscheduled hospital admissions in the United Kingdom.¹ Prompt and accurate diagnosis of the cause of these symptoms in crucial, particularly for conditions like acute myocardial infarction, acute heart failure or pulmonary embolism where delays in diagnosis and treatment may be associated with sudden death. Clinical assessment is not straightforward as symptoms and signs from these conditions are frequently indistinguishable from other common and less serious conditions. Key biomarkers are used to guide the diagnosis of acute cardiac conditions, but are often elevated in other conditions and can be challenging to interpret even for experienced clinicians.

Current pathways for the diagnosis of myocardial infarction, acute heart failure and pulmonary embolism have important limitations. First, the majority of care pathways rely on fixed biomarker thresholds that do not account for age, sex, or comorbidities, resulting in variable diagnostic performance and inequalities in care. Second, they are inefficient, with most patients requiring serial biomarker measurements or additional imaging with a low diagnostic yield. Third, with some notable exceptions, current care pathways do not incorporate important clinical features that have a major influence on pre-test probability. This results in diagnostic inaccuracies and inequalities in care. 

Artificial intelligence (AI) has potential to improve patient care by helping clinicians make more accurate and timely diagnoses. This project aims to create AI algorithms that combine biomarkers as continuous measures with routinely collected clinical data to calculate the diagnostic probability of multiple acute cardiovascular conditions in order to guide clinical decisions. Our previous AI algorithms have demonstrated superior performance compared to traditional methods for the diagnosis of myocardial infarction and heart failure.^2-4 However, current models are limited to single conditions and fail to account for the fact that patients often require evaluation for multiple conditions at once.

Aims

This project aims to develop and validate an AI-based decision-support tool for the Emergency Department that integrates multi-modal data to improve the diagnosis and classification of acute cardiovascular conditions. Specifically, the AI-based decision support tool will:

• Integrate diverse data sources by combining patient demographics, clinical history, cardiac biomarkers, and imaging data (e.g., ECG, chest X-rays) to improve diagnostic accuracy.
• Provide real-time diagnostic support by offering individualized diagnostic probabilities for multiple conditions, including myocardial infarction, heart failure, and pulmonary embolism, helping clinicians make faster and more precise decisions.
• Optimize resource use by improving efficiency, reducing unnecessary admissions, and enabling more targeted investigations based on personalized risk stratification.

We will use machine learning and deep learning approaches in consecutive patients with suspected acute cardiovascular disease from the Heart Disease Registry (https://dataloch.org/) as the training set to identify both informative and non-informative variables. Validation and testing will be conducted using patient data from Sweden, the United States, Australia, and New Zealand to ensure model calibration and generalisation in unseen datasets during the AI algorithm’s training process.

Training Outcomes

• Clinical and data science expertise: The student will gain experience in applying machine learning to real-world healthcare problems while developing a practical understanding of the problems at the interface of clinical practice and data analytics, including the language barrier with niche terminology on both ends. 
• Advanced statistical methodology and programming skills: The student will learn cutting-edge techniques in machine learning, deep learning, and developing interfaces which can be used by experts
• Interdisciplinary collaboration: Close work with clinicians and data scientists will enhance the student’s skills in interdisciplinary communication and translation of AI tools into clinical practice.

References

¹ British Heart Foundation Heart and Circulatory Diseases Statistics, British Heart Foundation, 2023. (www.bhf.org.uk).

² Doudesis D, Lee KK, Yang J, …, Mills NL. Validation of the myocardial-ischaemic-injury-index machine learning algorithm to guide the diagnosis of myocardial infarction in a heterogenous population: a prespecified exploratory analysis. The Lancet Digital Health 2022;4(5):e300-e308.

³ Lee KK, Doudesis D, Anwar M, …, Mills NL. Development and validation of a decision support tool for the diagnosis of acute heart failure: systematic review, meta-analysis, and modelling study. BMJ. 2022;377:e068424.

⁴ Doudesis D, Lee KK, Boeddinghaus J, …, Mills NL. Machine learning for diagnosis of myocardial infarction using cardiac troponin concentrations. Nature Medicine. 2023;29(5):1201-10.