Evaluating regression and probabilistic methods for ECG-based electrolyte prediction.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Medical Informatics

Background:

• Electrolyte imbalances pose significant health risks.
• Current blood-based electrolyte measurement is invasive, time-consuming, and not always accessible.
• Electrocardiograms (ECGs) are a rapid, non-invasive, and widely available diagnostic tool.

Purpose of the Study:

• To investigate the use of deep neural networks (DNNs) for predicting continuous electrolyte concentrations from ECG data.
• To compare DNN performance against traditional machine learning models for electrolyte prediction.
• To explore probabilistic regression for uncertainty estimation in ECG-based electrolyte monitoring.

Main Methods:

• Development and analysis of DNN models for regression tasks using a large dataset of over 290,000 ECGs.
• Comparison of DNNs with traditional machine learning algorithms.
• Evaluation of continuous prediction and binary classification of extreme electrolyte levels.
• Investigation of probabilistic regression and uncertainty quantification.

Main Results:

• DNNs demonstrated superior performance over traditional models in predicting electrolyte concentrations from ECGs.
• Model performance varied significantly across different electrolytes.
• Discretization achieved good classification accuracy but did not solve continuous prediction.
• Probabilistic regression showed potential, but uncertainty estimates require further calibration.

Conclusions:

• DNNs offer a promising, non-invasive approach for electrolyte concentration prediction using ECGs.
• Further development is needed to improve model generalizability across electrolytes and calibrate uncertainty estimates for clinical use.
• This research represents a foundational step towards a widely applicable ECG-based electrolyte monitoring system.