Improving diagnosis in acute cardiac care using statistical machine learning

Abstract

Cardiovascular disease affects more than half of people in the United Kingdom and remains the most common cause of death. Each year more than 25 million persons attend an Emergency Department, with chest pain or breathlessness being the most common presentations. These patients are often admitted to hospital because of concerns that they may have a life-threatening condition, such as acute myocardial infarction or decompensated heart failure. Despite the availability of specific and sensitive cardiac biomarkers, the diagnosis is not straightforward, resulting in unnecessary hospital admission or misdiagnosis. The aim of this thesis is to use cardiac biomarkers and statistical machine learning to develop clinical decision support tools that improve the diagnosis of patients presenting to the Emergency Department with possible acute cardiac conditions.

In 20,761 consecutive patients from the High Sensitivity Troponin in the Evaluation of Acute Coronary Syndrome (High-STEACS) trial, we validated a previously developed machine learning algorithm to assess its diagnostic performance for myocardial infarction in routine clinical practice. The myocardial-ischemic-injury-index (MI3) algorithm, which incorporates age, sex, and two troponin measurements of a patient, had excellent discrimination for the index diagnosis of myocardial infarction, and moreover, it predicted subsequent events too. However, the analysis showed that MI3 performance was not well calibrated in patients with intermediate probability, and there was considerable heterogeneity across important subgroups such as age, sex, presenting symptom of chest pain, cerebrovascular disease and renal function.

It is well known that cardiac troponin concentrations are influenced not only by the age and sex of the patient, but also by the time since symptom onset and comorbidities. Hence, we used patients from the High-STEACS trial and three external cohorts to develop and validate CoDE-ACS (Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome); a decision support tool that uses machine learning to incorporate clinical variables along with serial troponin measures and other laboratory tests. CoDE-ACS accurately predicts the likelihood of myocardial infarction and provides a more individualised diagnostic assessment. When CoDE-ACS was compared with the guideline-recommended clinical pathways, the performance was more consistent across important patient subgroups with better negative and positive predicted value.

We have subsequently developed and validated a decision support tool for patients with suspected acute heart failure; a condition where symptoms mimic many other conditions making the diagnosis challenging. To address this, we developed and externally validated CoDE-HF (Collaboration for the Diagnosis and Evaluation of Heart Failure) in 10,369 patients from 13 countries to improve the diagnosis and evaluation of acute heart failure. CoDE-HF combines blood natriuretic peptide concentrations as a continuous measure and simple objective clinical variables known to be associated with acute heart failure. First, we used the N-terminal pro-B-type natriuretic peptide (NT-proBNP), as it is the most common test used in clinical practice. Then, we retrained CoDE-HF to support the use of two other natriuretic peptides (B-type natriuretic peptide [BNP] and mid-regional pro atrial natriuretic peptide [MR-proANP]). Last, we compared my solution with the guideline-recommended approach showing that my new decision support tool could achieve a better overall performance, including in complex patients with comorbidities.

My findings suggest that a precision medicine approach combining machine learning algorithms with clinical variables and cardiac biomarkers could improve the diagnostic information provided to clinicians when assessing patients with suspected myocardial infarction and heart failure in the Emergency Department.