ECG-Based Detection of Acute Myocardial Infarction Using a Wrist-Worn Device.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Wrist-worn electrocardiography (wECG) devices offer a potential method for acute myocardial infarction (AMI) detection.
• The sufficiency of wECG data for accurate AMI diagnosis remains under investigation.

Purpose of the Study:

• To investigate explainable machine learning models for AMI detection using wECG data.
• To compare the performance of convolutional neural networks (CNNs) and gradient-boosting decision trees (GBDTs) for wECG-based AMI detection.

Main Methods:

• Two machine learning models, CNN (raw ECG input) and GBDT (feature input), were developed.
• 123 participants (AMI patients, other cardiovascular disease patients, healthy controls) were enrolled.
• A wrist-worn device acquired limb lead I and a chest/abdominal lead (V3, V5, or abdomen).

Main Results:

• Models incorporating all four leads demonstrated optimal performance.
• CNN achieved a sensitivity of 0.77 and specificity of 0.75; GBDT achieved 0.77 sensitivity and 0.72 specificity.
• CNN and GBDT showed high specificity (0.94 and 0.90, respectively) when differentiating AMI from healthy individuals.

Conclusions:

• wECG-based AMI detection holds significant potential for out-of-hospital applications.
• Explanations from CNN models showed limited agreement with traditional ECG interval analysis, necessitating careful clinical interpretation.