Machine Learning Prediction of Blood Potassium at Different Time Cutoffs.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Serum potassium levels significantly impact cardiac function and ECG morphology.
• The temporal relationship between ECG changes and potassium levels is crucial for accurate prediction.
• Machine learning (ML) offers a promising approach for non-invasive potassium level estimation using ECG data.

Purpose of the Study:

• To investigate the sensitivity of ML-based serum potassium prediction using 12-lead ECG data.
• To evaluate the impact of the time interval between ECG acquisition and serum potassium measurement on prediction accuracy.
• To assess model performance under different data availability scenarios.

Main Methods:

• A convolutional neural network (CNN) was trained to classify normal versus abnormal serum potassium levels (≥5 mEq/L) using only ECG data.
• Model performance was compared using data filtered within 1-hour, 30-minute, and 15-minute intervals.
• Scenarios included using all available data and restricted datasets to match training set sizes.

Main Results:

• The 1-hour interval with all data yielded the highest average AUC (0.850) and weighted accuracy (76.3%).
• The 15-minute interval resulted in an average AUC of 0.814 and weighted accuracy of 72.5%.
• Restricting training set sizes to match the 15-minute cutoff yielded comparable AUC and accuracy across all time intervals.

Conclusions:

• The time interval between ECG and serum potassium measurement influences ML prediction performance.
• ML models demonstrate potential for predicting serum potassium levels from ECG, even with varying temporal data.
• Future research should focus on failure case analysis, bias identification, and explainability for robust ML potassium prediction.