Machine learning compared with rule-in/rule-out algorithms and logistic regression to predict acute myocardial infarction based on troponin T concentrations | JoVE Visualize

Machine learning compared with rule-in/rule-out algorithms and logistic regression to predict acute myocardial infarction based on troponin T concentrations

Anders Björkelund¹, Mattias Ohlsson¹, Jakob Lundager Forberg²

¹Department of Astronomy and Theoretical Physics Lund University Lund Sweden.

Journal of the American College of Emergency Physicians Open|March 29, 2021

View abstract on PubMed

Summary

Machine learning algorithms, like artificial neural networks (ANNs), can improve the diagnosis of acute myocardial infarction (AMI) in emergency departments. These tools enhance risk assessment for chest pain patients by reducing the uncertain intermediate group.

Area of Science:

Cardiology
Medical Informatics
Artificial Intelligence

Background:

Acute myocardial infarction (AMI) diagnosis in emergency departments (ED) relies on clinical assessment and biomarkers.
Current diagnostic algorithms for AMI using high-sensitivity cardiac troponin T (hs-cTnT) have limitations regarding sampling times.
Computerized decision-support tools offer potential to refine AMI diagnosis.

Purpose of the Study:

To evaluate the predictive accuracy of machine learning algorithms for ruling in or out AMI.
To compare the performance of an artificial neural network (ANN) against standard European guideline algorithms for hs-cTnT.
To assess the impact of varying sampling times on diagnostic performance for AMI.

Main Methods:

Retrospective, register-based diagnostic study of 5695 chest pain patients from two Swedish hospitals (2013-2014).
Utilized 5-fold cross-validation 200 times to compare ANN performance with 0/1- and 0/3-hour hs-cTnT algorithms and logistic regression.
Primary outcome: size of the intermediate risk group (AMI uncertain) with constant sensitivity and specificity.

Main Results:

ANN and logistic regression demonstrated high predictive accuracy (95% area under the ROC curve).
In patients meeting sampling time criteria (n=4171), ANN reduced the intermediate risk group by 9.2% compared to guideline algorithms (24.5% to 22.2%).
Logistic regression showed no significant reduction in the intermediate group size.

Conclusions:

Machine learning algorithms, particularly ANNs, offer enhanced flexibility in hs-cTnT sampling for AMI diagnosis.
These algorithms show potential to improve risk stratification for chest pain patients in the ED.
Machine learning may lead to more efficient and accurate diagnosis of acute myocardial infarction.