A generalizable electrocardiogram-based artificial intelligence model for 10-year heart failure risk prediction.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Heart failure (HF) is a global health concern with two primary subtypes: HFpEF and HFrEF.
• Early risk prediction and modification are crucial for managing HF progression.
• There is a need for accessible, AI-driven tools for early HF risk assessment.

Purpose of the Study:

• To validate artificial intelligence (AI) models for heart failure (HF) risk prediction using Multi-Ethnic Study of Atherosclerosis (MESA) data.
• To assess the performance of various models, including ECG-AI, in classifying HFpEF and HFrEF.
• To compare the predictive accuracy of AI-based models against traditional clinical risk calculators.

Main Methods:

• Six models were compared: an ECG-AI model (convolutional neural network), clinical models (ARIC-HF, FHS-HF), Cox proportional hazards (CPH) models (CPH, ECG-AI-Cox), and an ECG Characteristics (ECG-Chars) model.
• Models were trained using ARIC data and validated on MESA data.
• Performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC) and compared via the DeLong test.

Main Results:

• The ECG-AI-Cox model achieved the highest validation AUC (0.84), outperforming other models.
• Specific AUCs were: ECG-AI (0.77), ECG-Chars (0.73), ARIC-HF (0.76), FHS-HF (0.74), CPH (0.78).
• ECG-AI-Cox showed strong performance in classifying HFrEF (AUC=0.85) and HFpEF (AUC=0.83).

Conclusions:

• AI models utilizing ECG data offer superior validated predictions compared to traditional HF risk calculators.
• The ECG-AI approach, particularly the ECG-AI-Cox model, effectively predicts HF risk and aids in HFpEF and HFrEF classification.
• These findings support the use of AI-powered ECG analysis for early and accurate heart failure detection.