Development and validation of a dynamic deep learning algorithm using electrocardiogram to predict dyskalaemias in...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Deep learning models (DLMs) excel in electrocardiogram (ECG) analysis for diagnosing electrolyte imbalances like dyskalemia.
• Previous research has not evaluated DLM performance in continuous patient follow-up scenarios.
• Accurate, continuous monitoring is crucial for managing patients with fluctuating potassium levels.

Purpose of the Study:

• To develop and validate a dynamic revision of DLM-enabled ECG analysis for improved accuracy in diagnosing dyskalemia.
• To enhance diagnostic precision in patients requiring multiple ECG assessments over time.
• To assess the applicability of this revised method in other clinical predictions, such as ejection fraction.

Main Methods:

• Retrospective collection of 168,450 ECGs and corresponding serum potassium levels from 103,091 patients for model development.
• Internal and external validation using 37,246 and 47,604 ECGs, respectively.
• Implementation of a dynamic revision method utilizing personal pre-annotated ECGs for continuous follow-up analysis.

Main Results:

• The dynamic revision method demonstrated superior performance compared to traditional direct prediction for diagnosing hypokalemia (AUC 0.730/0.720-0.788/0.778) and hyperkalemia (AUC 0.884/0.888-0.915/0.908).
• The model achieved significant accuracy in both internal and external validation sets.
• The method showed a distinguishable improvement in diagnosing dyskalemia in patients with multiple visits.

Conclusions:

• The proposed dynamic revision significantly enhances the accuracy of DLMs for diagnosing dyskalemia in patients with multiple visits.
• This approach offers a valuable tool for continuous ECG monitoring and management of electrolyte imbalances.
• The method's successful application in ejection fraction prediction suggests broader clinical utility.