Artificial Intelligence-Enabled Electrocardiography for Monitoring Serum Potassium Dynamics in Patients With Severe...

Severe hypokalemia requires prompt management and close surveillance. Although artificial intelligence-enabled electrocardiography (AI-ECG) rapidly detects severe hypokalemia, its application for monitoring serum potassium (K⁺) dynamics during treatment remains unexplored. This study assessed AI-ECG performance in monitoring K⁺ changes during supplementation.

Multicenter retrospective cohort study.

191 adults with severe hypokalemia (laboratory-measured [lab-K⁺] ≤ 2.5 mmol/L; matched ECG-K⁺ < 3.5 mmol/L) and ≥1 follow-up paired measurement within 24 hours of K⁺ supplementation at 3 teaching hospitals between September 2019 and August 2024.

Laboratory-measured K⁺ (lab-K⁺) and K⁺ estimated by ECG (ECG-K⁺) overall and stratified by the etiology of hypokalemia (acute K⁺ shift vs chronic K⁺ deficit).

Primary: agreement between paired ECG-K⁺ and lab-K⁺. Secondary: diagnostic accuracy and K⁺ trajectories.

Linear mixed-effects models with patient-level random intercepts; repeated-measures correlation (rmcorr) and Bland-Altman plots; patient-level clustered bootstrapped receiver operating characteristic analysis for diagnostic accuracy.

Of 191 patients, 156 (81.7%) had chronic K⁺ deficits (most commonly gastrointestinal disorders [n = 47] or diuretic use [n = 35]), and 35 (18.3%) had acute K⁺ shifts (most commonly thyrotoxic periodic paralysis [n = 25]). The chronic K⁺ deficits group had more comorbidities and use of medications affecting K⁺. ECG-K⁺ correlated strongly with lab-K⁺ (rmcorr 0.847 [95% CI, 0.81-0.88], P < 0.001). The relationship was modified by hypokalemia etiology (interaction P < 0.0001) with a lower correlation in patients with chronic K⁺ deficits. The diagnostic accuracy of ECG-K⁺ with lab-K⁺ < 3.5 mmol/L was reflected by an area under the curve of 0.920 (95% CI, 0.863-0.961). It was higher in patients with an acute K⁺ shift. ECG-K⁺ preceded lab-K⁺ results by a mean of 52.5 minutes. Patients with an acute K⁺ shift corrected approximately 3-fold faster than those with a chronic K⁺ deficit (0.121 vs 0.039 mmol/L/h). Rebound hyperkalemia was detected by ECG-K⁺ in 2 patients before laboratory confirmation.

Retrospective design, treatment-protocol heterogeneity, limited inpatient medication granularity, and potential selection bias.

AI-ECG enables real-time, within-patient monitoring of serum K⁺ dynamics during treatment for severe hypokalemia, with superior performance in the setting of acute hypokalemia due to K⁺ shift. As a noninvasive adjunct, AI-ECG may shorten the time to detect changes in K⁺ and reduce the need for lab-K⁺ measurements rather than relying exclusively on lab-K⁺. Confirmatory studies are warranted.

Severe hypokalemia, or very low blood potassium, is common in hospitalized patients and can cause dangerous heart rhythms. In clinical practice, potassium levels are often monitored with blood tests whose results are not immediately available. This study assessed whether an artificial intelligence (AI) tool that estimates blood potassium levels from a standard electrocardiogram (ECG) tracing can reduce the need for laboratory testing of blood for potassium levels. Patients with hypokalemia at 3 hospitals were studied. AI-based estimates of blood potassium were obtained very close in time to when blood was drawn for laboratory measurements of potassium. The AI-based estimates were available almost an hour earlier than laboratory tests, and both values were very similar. The similarity was greater when the hypokalemia was caused by a sudden shift into cells rather than chronic potassium loss from the body. AI-based estimates of potassium potentially offer a faster, noninvasive way to track potassium changes for patients over time, to guide potassium replacement, to reduce the need for repeated blood draws, and to detect dangerously high levels of potassium earlier.