Identifying Ventricular Dysfunction Indicators in Electrocardiograms via Artificial Intelligence-Driven Analysis

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Artificial intelligence (AI) shows promise in identifying ventricular dysfunction from electrocardiograms (ECGs).
• Specific ECG waveforms indicative of cardiac dysfunction remain largely undefined.
• Accurate, non-invasive methods for assessing ventricular function are crucial in clinical practice.

Purpose of the Study:

• To develop and validate AI models for detecting reduced left ventricular ejection fraction (LVEF) using ECG data.
• To identify specific ECG configurations and waveform segments most effective for diagnosing ventricular dysfunction.
• To determine the diagnostic utility of different ECG leads and segments for assessing cardiac function.

Main Methods:

• Analysis of ECG and echocardiography data from 17,422 patients in Japan and Germany.
• Development of 10-layer convolutional neural networks (CNNs) for LVEF classification (<50%).
• Evaluation of model performance using four-fold cross-validation across various ECG configurations (3s strips, single-beat, two-beat overlay) and segments (PQRST, QRST, P, QRS, PQRS).

Main Results:

• Two-beat ECG configurations demonstrated superior performance in detecting ventricular dysfunction compared to single-beat and 3s strip models.
• Single-beat models identified limb leads I and aVR as particularly indicative of dysfunction.
• ECG segments from the QRS complex to the T-wave were most informative, with P-wave segments further enhancing diagnostic accuracy.

Conclusions:

• Dual-beat ECG analysis, processed by AI, enables highly precise classification of ventricular function.
• Specific ECG segments, notably from the P-wave through the T-wave, are more effective for assessing ventricular dysfunction.
• ECG leads I and aVR exhibit significant diagnostic utility for evaluating ventricular dysfunction.