Identifying heart failure dynamics using multi-point electrocardiograms and deep learning.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Heart failure (HF) hospitalizations are linked to poor survival, necessitating early intervention.
• Deep learning (DL) shows promise for HF detection via electrocardiograms (ECGs), but its role in ongoing monitoring is unclear.
• Current monitoring methods may be invasive or costly, highlighting a need for accessible alternatives.

Purpose of the Study:

• To develop and evaluate a DL model for detecting changes in HF status using serial ECGs.
• To enhance early intervention and continuous HF monitoring capabilities in diverse healthcare settings.
• To assess the model's performance in classifying HF status changes (deteriorated, improved, no-change).

Main Methods:

• A Transformer-based DL model was developed using 30,171 ECGs from 6,531 adult patients.
• ECGs were collected at two different time points to analyze HF status changes.
• Model performance was evaluated using AUROC and accuracy; attention mapping was used for interpretability.

Main Results:

• The DL model achieved an AUROC of 0.889 and an accuracy of 0.871 for HF status classification.
• The model effectively identified changes in HF status based on ECG waveform signals.
• Attention mapping provided insights into the model's decision-making process.

Conclusions:

• The Transformer-based DL model accurately detects HF status changes, offering a non-invasive monitoring tool.
• This approach reduces the need for invasive and costly diagnostic procedures.
• The model holds potential for accessible and efficient HF management, supporting early intervention.