Get started
Work

Artificial intelligence and machine learning to improve the diagnosis of acute myocardial infarction

Year: 2026
Type: dissertation
Abstract: INTRODUCTION: Myocardial infarction is a leading cause of death and morbidity worldwide. In the United Kingdom alone, patients with possible myocardial account for around one in ten of the 30 million visits to the Emergency Department each year. An early triage decision is needed in these patients in order to identify those at highest risk who most likely benefit from immediate treatment. On the other hand, most patients do finally not have an acute myocardial infarction, and it is equally important to identify these patients, as the majority are eligible for early discharge. To guide clinicians, accelerated diagnostic pathways using cardiac troponin measurements have been established. These pathways are safe and effective and recommended by international practice guidelines. However, several challenges remain. These include differences in cardiac troponin assays used in clinical practice, limitations of the diagnostic pathways applied, and difficulties in assessing the cardiovascular risk of patients with myocardial injury or infarction. The aim of this thesis is to address these challenges in the diagnosis of myocardial infarction and risk assessment of patients by using novel tools that incorporate artificial intelligence and machine learning to aid clinical decision making. METHODS: Study 1 was a prospective interrupted time-series analysis. In this study, consecutive patients presenting with suspected acute coronary syndrome to one of three acute care hospitals in Scotland were identified. All sites changed from an assay measuring cardiac troponin I to one measuring cardiac troponin T. The primary outcome was hospital admission, and secondary outcomes included subsequent myocardial infarction, heart failure hospitalisation or cardiovascular death at one year. Study 2 was an individual patient-level data meta-analysis of prospective studies where cardiac troponin was measured for suspected acute coronary syndrome, diagnoses were adjudicated using the Universal Definition, and both myocardial and cause-specific mortality was reported at a minimum of one year. The primary outcome was major adverse cardiovascular events defined as myocardial infarction or cardiovascular death at one year. Study 3 was a derivation and validation study. In this study, machine learning models were trained to integrate cardiac troponin at presentation or on serial testing with clinical features to compute the CoDE-ACS score (0-100) corresponding to an individual's probability of myocardial infarction. The model was trained and externally validated to identify patients with an adjudicated diagnosis of type 1, type 4b or type 4c myocardial infarction during the index hospital admission and tested across various subgroups, countries and healthcare systems. Study 4 was a secondary analysis of an international prospective cohort study. Patients with possible myocardial infarction were enrolled and underwent serial high-sensitivity cardiac troponin I measurements at 0, 1 and 2 hours. The primary outcome was the diagnostic performance of the CoDE-ACS model at each timepoint for index type 1 myocardial infarction and the efficacy of previously validated low- and high-probability scores compared with guideline-recommended pathways. RESULTS: In study 1, 25,849 patients were enrolled. Of these, 13,146 (median age 60 [interquartile range 47 to 73] years, 47% female) and 12,703 (60 [47 to 73] years, 46% female) presented before and after the transition to cardiac troponin T, respectively. The proportion of patients with myocardial injury increased from 21% (2,800/13,146) to 38% (4,781/12,703) and patients were more likely to be admitted to hospital following the transition (odds ratio 2.24, 95% confidence interval, 1.81 to 2.77, both P<0.001). A 6-fold increase in serial testing was observed in patients undergoing a measure of cardiac troponin T compared to cardiac troponin I (odds ratio 6.03, 95% confidence interval, 4.85 to 7.49, P<0.001). Subsequent myocardial infarction, heart failure or cardiovascular death at one year was comparable before and after the transition (odds ratio 0.83, 95% confidence interval, 0.48 to 1.41, P=0.487). In study 2, 17 studies of nine cohorts from nine countries, with individual patient-level data available in eight cohorts, were identified. Among 120,734 patients (median 61.0 years, 45.8% female), type 1 myocardial infarction occurred in 9.4% (n=11,298), type 2 myocardial infarction in 3.0% (n=3,609), acute myocardial injury in 4.9% (n=5,864) and chronic myocardial injury in 4.7% (n=5,625). In patients with type 1 and type 2 myocardial infarction, the major adverse cardiovascular event rate was 55.2 and 51.7 per 1,000 patient-years, and the non-cardiovascular death rate was 25.7 and 60.1 per 1,000 patient-years, respectively. Compared to those without myocardial injury, the risk of major adverse cardiovascular events (subdistribution hazard ratio) was 4.82 (95% confidence interval, 3.55–6.57, I²=84%), and 3.36 (95% confidence interval, 2.92–3.86, I²=0%), respectively. In patients with acute and chronic myocardial injury, the major adverse cardiovascular event rate was 47.1 and 44.9 per 1,000 patient-years, and the non-cardiovascular death rate was 67.0 and 46.9 per 1,000 patient-years, respectively. The risk of major adverse cardiovascular events was 3.24 (95% confidence interval, 2.41-4.36, I2=59%) and 3.03 (95% confidence interval, 2.53-3.62, I2=20%), respectively. In study 3, the derivation cohort comprised of 10,038 patients (median age 70 years, 48% women) and the external validation cohort consisted of 10,286 patients (median age 60 years, 35% women) with suspected myocardial infarction. Discrimination of the CoDE-ACS models was excellent with an area under curve of 0.953 (95% confidence interval, 0.947, 0.958) at presentation and 0.966 (95% confidence interval, 0.961, 0.970) on serial testing. In the external validation cohort, a CoDE-ACS score of less than 3 identified 61% (6,265/10,286) of patients as low probability at presentation with a negative predictive value of 99.6 (95% confidence interval, 99.4, 99.7) and a sensitivity of 97.9 (95% confidence interval, 97.6, 98.2). A CoDE-ACS score of 61 or greater identified 10% (1,052 /10,286) of patients at presentation as high probability with a positive predictive value of 75.5 (95% confidence interval, 74.6, 76.3) and a specificity of 97.1 (95% confidence interval, 96.8, 97.4). Compared to patients identified by CoDE-ACS at presentation as intermediate- or high-probability, those who were low-probability of myocardial infarction had a lower rate of cardiac death and all-cause death at 30-days (cardiac death: 0.1% versus 0.5% and 1.8%; all-cause death: 0.1% versus 0.9% and 2.0%, respectively) and at one year (cardiac death: 0.3% versus 2.8% and 4.2%; all-cause death: 1.1% versus 6.1% and 6.7%, respectively, log-rank test P<0.001) In study 4, a total of 4,105 patients (age 61 [50-74] years, 32% women) were included where 575 (14%) had type 1 myocardial infarction. At presentation, CoDE-ACS identified 56% of patients as low-probability, with a negative predictive value and sensitivity of 99.7% (95% confidence interval, 99.5-99.9%) and 99.0% (95% confidence interval, 98.6-99.2%), ruling out more patients than the European Society of Cardiology 0h and High-STEACS (25% and 35%) pathways. CoDE-ACS incorporating a second cTn measurement identified 65% or 68% of patients as low probability at 1 or 2 hours, for an identical negative predictive value of 99.7% (95% confidence interval, 99.5-99.9%), 19% or 18% as high-probability with a positive predictive value of 64.9% (95% confidence interval, 63.5-66.4%) and 68.8% (95% confidence interval, 67.3-70.1%), and 16% or 14% as intermediate probability. In comparison, the European Society of Cardiology 0/1h, European Society of Cardiology 0/2h and High-STEACS pathways after serial measurements identified 49%, 53% and 71% of patients as low-risk with a negative predictive value of 100% (95% confidence interval, 99.9-100%), 100% (95% confidence interval, 99.9-100%) and 99.7% (95% confidence interval, 99.5-99.8%), and 20%, 19% or 29% as high-risk with a positive predictive value of 61.5% (95% confidence interval, 60.0-63.0%), 65.8% (95% confidence interval, 64.3-67.2%), and 48.3% (95% confidence interval, 46.8-49.8%) resulting in 31%, 28% or 0% who require further observation in the Emergency Department, respectively. CONCLUSION: Challenges in the early risk stratification of patients with suspected myocardial infarction and the risk assessment of those with confirmed myocardial injury or infarction persist. These challenges can be addressed using novel clinical decision-support tools incorporating artificial intelligence and machine learning. Such a personalised approach allows to assess an individual's risk of having an acute myocardial infarction. Future research is needed to determine whether implementation of clinical decision-support tools using artificial intelligence and machine learning will improve patient care, reduce overcrowding in the Emergency Department, and saves healthcare costs.
DOI: 10.7488/era/8011
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7225915320
Cites:
Cited by:
Related to: 10