Predicting Serum Potassium from ECG Using Deep Neural Networks: A Pilot Study on a New Open-Access Dataset

Dyskalemia, encompassing hyperkalaemia (K⁺ ≥ 5.5 mEq/L) and hypokalaemia (K⁺ ≤ 3.5 mEq/L), is a frequent electrolyte disorder with potentially lead to fatal cardiac consequences. While diagnosis depends on invasive and time-consuming blood tests, the influence of serum K⁺ on ECG morphology enables ECG-based assessment. To overcome the limited generalisability of prior DNN approaches based on proprietary data, an open benchmark dataset using MIMIC-IV data is introduced in this study. Ten-second 12-lead ECG recordings were paired with serum K+ measurements taken within a one-hour window, resulting in 162,318 samples from 74,488 patients. Signals were bandpass filtered (0.5-100 Hz), z-normalized, and divided via patient-level split into training (80%), validation (10%), and test (10%) sets. Multiple DNN architectures, including 1D ResNet and CNN-Transformer (Conformer) with multimodal fusion, were trained to distinguish healthy from dyskalemic patients. Performance was evaluated using the area under the receiver operating characteristics curve (AUC). Best performance was achieved with a Conformer architecture (1D CNN + Transformer), yielding AUCs of 0.795 for hypokalaemia and 0.724 for hyperkalaemia in binary classification. In a cost-sensitive 7-class severity task, performance remained strong for severe hypokalaemia (AUC 0.769 for K⁺ < 2.5 mEq/L) but declined for severe hyperkalaemia (AUC 0.624 for K⁺ > 6.5 mEq/L), likely due to potassium-specific label noise such as haemolysis and timestamp inaccuracies. Ongoing work focuses on improved data curation, subgroup analyses, and external validation. This study provides preliminary evidence that ECG-based deep learning can support non-invasive dyskalaemia screening, with reliable hypokalaemia detection and ongoing efforts to resolve current limitations in hyperkalaemia prediction through dataset refinement and external validation.