Development of a decision support tool to improve the diagnosis and classification of myocardial infarction using signal processing and statistical machine learning

Background

Our research aims to harness routine electronic patient data to improve the diagnosis and targeting of treatment for patients with coronary heart disease. We have linked multiple routinely collected electronic data sources through the network of NHS safe havens to evaluate the impact of new approaches for diagnosis and risk assessment in consecutive patients with suspected acute coronary syndrome across Scotland in a series of stepped-wedge cluster randomized clinical trials. We have demonstrated that (a) the introduction of a more sensitive cardiac troponin test into clinical practice was associated with improvement in survival of patients due to the better targeting of treatments for coronary heart disease [1]; (b) contemporary troponin tests underdiagnose myocardial infarction in women and contribute to sex-inequalities in treatment [2]; (c) high-sensitivity tests can be used to identify low risk patients at presentation who may not require hospital admission [3]. These observations have changed myocardial infarction assessment [4], and are changing national and international guidelines.
Currently, diagnosis is based on a fixed threshold of cardiac troponin, and does not adequately take into consideration important individual patient characteristics, such as age, sex, or comorbidities, or take into account any interaction of these variables with one another. Therefore, in a large, international, multi-center study of patients with suspected myocardial infarction, we applied a supervised learning approach to train and test a decision tree (the myocardial ischemic injury index, MI3) that incorporates simple, objective variables to accurately predict the likelihood of a diagnosis of myocardial infarction [manuscript under review]. MI3 incorporates age, sex, high-sensitivity cardiac troponin I concentrations at presentation and 1-3 hours later, and the rate of change in cardiac troponin I concentrations between testing. The algorithm returns an outcome in the range 0-100 reflecting the probability of acute myocardial infarction. The algorithm was well calibrated and diagnostic performance was very good across large training and testing sets including more than 11,000 patients and 1,250 patients with myocardial infarction.

Encouraged by these results, we want to extend our work to include new data mining tools to process the electrocardiogram and fuse multi-modal information from additional clinical tests. These approaches will provide guidance on the likely aetiology and classification of the diagnosis, assisting clinical decisions in practice where diagnosis is less certain and patients often present with multiple comorbid conditions simultaneously. We hope to evaluate the implementation of this tool into clinical practice, to determine whether providing clinicians with diagnostic probabilities for individual patients, will improve the targeting of further investigation and treatments for myocardial infarction.

Aims
The aim of the project is to develop a new clinical decision support tool that will provide accurate probabilistic classification of myocardial infarction for the evaluation of patients with acute chest pain in the Emergency Department. We will use the HighSTEACS trial (n=54,000, NCT:01852523) as the training set and the HiSTORIC trial (n=34,000, NCT03005158) as the testing set to report performance of the machine learning algorithm. In both trials the labeling (type 1-5 myocardial infarction) was performed by a team of physicians using an established adjudication portal and web-interface (Figure). The project will develop new data mining tools to process the electrocardiograms and fuse multi-modal information from additional clinical tests to classify MI types.

Training outcomes
• Practical understanding of the problems at the interface of clinical practice and data analytics, including the language barrier with niche terminology on both ends
• Developing expertise in time-series analysis, signal processing, and statistical machine learning in order to tackle large-scale challenging problems
• Programming skills: transforming algorithmic concepts to software tools, and developing interfaces which can be used by experts

This MRC programme is joint between the Universities of Edinburgh and Glasgow. You will be registered at the host institution of the primary supervisor detailed in your project selection.

All applications should be made via the University of Edinburgh, irrespective of project location:

http://www.ed.ac.uk/studying/postgraduate/degrees/index.php?r=site/view&id=919

Please note, you must apply to one of the projects and you are encouraged to contact the primary supervisor prior to making your application. Additional information on the application process if available from the link above.

For more information about Precision Medicine visit:

http://www.ed.ac.uk/usher/precision-medicine