Clinical decision support using machine learning and cardiac troponin for the diagnosis of myocardial infarction

Current assessment of chest pain and possible myocardial infarction is based on cardiac troponin (cTn) measurement, electrocardiogram (ECG), and either structured or subjective risk assessment. Pathway performance varies with patient age, sex, and ethnicity and the time of blood sampling from symptom onset and between samples. This heterogeneity is not accounted for in strategies using fixed cTn thresholds for all patients or in those that categorize patients into two to four risk groups. This compartmentalization at best crudely recognizes risk differences due to demographics or small variation in cTn results or timing. It requires considerable clinical experience for the human brain to accurately weigh all of these factors when making decision. Machine learning (ML) is a quantitative and reproducible way to combine multiple variables to improve predictive accuracy. There are multiple ML methodologies available of which most are variations of statistical regression modelling. Generally, they allow for individualized risk assessment by combining variables, both quantitative and categorical. They account for non-linear complex interactions between variables. Machine learning can also incorporate temporal variation such as the rate of change in troponin concentration. This is important because contemporary pathways rely on fixed absolute or relative change in cTn thresholds, which require consistency in the timing of blood sampling that is not always possible in a busy Emergency Department. In contrast, ML can account for variation in the time intervals between blood draws. Three ML approaches that estimate the probability of myocardial infarction in the Emergency Department have been derived and validated: Myocardial-Ischaemic-Injury-Index (MI3), Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome (CoDE-ACS), and ARTificial intelligencE in suspected Myocardial Infarction Study (ARTEMIS).1–3 The MI3 incorporates age, sex, and paired cTnI measurements over 1–3 h. A low MI3 score identified 69.5% of patients as low probability after serial testing with a negative predictive value (NPV) of 99.7% and sensitivity of 97.8%.1 A high MI3 score identified 10.6% of patients as high probability, with a positive predictive value (PPV) of 71.7% and specificity of 96.7%. The MI3 score performed better than a conventional 0/3 h cTn pathway using the 99th percentile. The CoDE-ACS can be calculated using a single cTn measurement and incorporates additional variables including age, sex, comorbidities, previous ischaemic heart disease, chest pain, time from symptom onset, ECG, blood pressure, and pulse rate.2 Following a single cTn measurement, a low CoDE-ACS score identified twice as many patients as low probability as a low cTn threshold (61% vs. 27%) with a similar NPV (99.6% vs. 99.7%) and sensitivity (98.3% vs. 97.9%). A high CoDE-ACS score was superior to the use of sex-specific 99th percentile with PPVs of 85.1% and 63.6%, respectively. The CoDE-ACS with a single or serial cTn measurement was more effective than the current guideline recommended pathways.3 The ARTEMIS was trained using multiple assays and combines age, sex, symptom onset time, ECG, and several other cardiovascular risk factors with cTn.4 The probability of myocardial infarction was calculated with high accuracy, achieving a PPV > 70% for all but one assay. A good clinical decision aid is easy to use, fits well within the clinician’s workflow, and saves time. Unfortunately, user interface and, consequently, experience are often poor. Key history and presentation variables are needed for accurate prediction and should be, but rarely are, incorporated into the medical record. Wayfind is an example of such a decision support tool (Figure 1).5 It provides numerical and visual depiction using an icon array of probabilities calculated using local pathways. Depending on local acceptability and resource availability, it can provide specific patient management guidance tailored to the institution. Workload is reduced by providing direct links to further tasks (e.g. ordering further tests or printing information for a patient). Importantly, the tool will output an easy-to-read medical record, which makes the patient management decision less burdensome. Illustration of the retraining of machine learning models to support clinical decisions in patients with possible myocardial infarction. Visual representation depicts sequential data capture of key variables in the assessment of a patient with possible myocardial infarction. Overlapping windows depict user interfaces for patient demographics, presenting problem and past medical history, symptoms and signs, and the electrocardiogram and cardiac troponin results. A machine learning algorithm estimates diagnostic probabilities in the background and presents an individualized probability of myocardial infarction both numerically and visually (using an icon array display). Recommended actions for the clinician are given based on this probability. Once risk predictions are linked to data on patient outcomes, the model can be recalibrated to improve accuracy. The combination of routine high-fidelity data capture matched to prediction, outcomes, and recalibration allows for the creation of a learning healthcare system. Images from Wayfind Health, used with permission. Ideally, ML algorithms would undergo verification, particularly calibration, in each hospital before implementation.6 This will be challenging for most hospitals, but decision support tools that collect the relevant data could achieve this if data are linked to clinical outcomes. This is the basis of a learning healthcare system that can evolve in time, accounting for changing demographics of patients, which has major potential to improve both effectiveness and safety of decision-making for patients with possible myocardial infarction (Figure 1). None. Members of the Study Group on Biomarkers of the Association for Acute CardioVascular Care of the ESC: Bertil Lindahl, MD1; Jasper Boeddinghaus, MD2; Louise Cullen, MD, PhD3; Lori Daniels, MD4; Ola Hammarsten, MD, PhD5; Kurt Huber, MD6; Evangelos Giannitsis, MD7; Allan S. Jaffe, MD8; Dorien M. Kimenai, PhD9; Konstantin A. Krychtiuk, MD, PhD10; Martin Möckel, MD11; Christian Mueller, MD2; Matthias Thielmann, MD12; Kristian Thygesen, MD13; Johannes Mair, MD14; Nicholas L. Mills, MD, PhD9,15 1Department of Medical Sciences, University of Uppsala, Uppsala, Sweden 2Cardiovascular Research Institute Basel (CRIB) and Department of Cardiology, University Hospital Basel, University of Basel, Switzerland 3Emergency and Trauma Centre, Royal Brisbane and Women’s Hospital, Brisbane, Australia 4Department of Medicine, Sulpizio Cardiovascular Center, University of California, San Diego, La Jolla, CA, USA 5Department of Clinical Chemistry and Transfusion Medicine, University of Gothenburg, Gothenburg, Sweden 6Department of Medicine, Cardiology and Intensive Care Medicine, Wilhelminen Hospital and Sigmund Freud University, Medical School, Vienna, Austria 7Department of Cardiology, University Heidelberg, Germany 8Departments of Cardiology and Laboratory Medicine and Pathology, Mayo Clinic, Rochester, Minnesota, USA 9BHF Centre for Cardiovascular Science, University of Edinburgh, Edinburgh, UK 10Department of Internal Medicine II, Division of Cardiology, Medical University of Vienna, Vienna, Austria 11Division of Emergency Medicine, Charité-Universitätsmedizin Berlin, Germany 12Klinik für Thorax-und Kardiovaskuläre Chirurgie, Westdeutsches Herz- und Gefäßzentrum Essen, University of Duisburg-Essen, Germany 13Department of Cardiology, Aarhus University Hospital, Denmark 14Department of Internal Medicine III—Cardiology and Angiology, Innsbruck Medical University, Innsbruck, Austria 15Usher Institute, University of Edinburgh, Edinburgh, UK