Introduction and background

Hyperkalemia and hypokalemia are common electrolyte disturbances encountered across emergency, inpatient, intensive care, dialysis, and ambulatory settings. They arise from kidney disease, endocrine disorders, acute illnesses, or medications. Severe potassium abnormalities can precipitate life-threatening arrhythmias requiring urgent treatment [1]. Timely recognition is therefore essential. Laboratory measurement of serum, plasma, or whole-blood potassium remains the diagnostic reference standard and continues to guide treatment decisions [1].

The electrocardiogram (ECG) has long been used to look for potassium-related changes, but its conventional findings are an unreliable substitute for laboratory potassium measurement. Classic features such as peaked T waves or QRS widening are inconsistently present and are influenced by baseline cardiac disease, medications, conduction abnormalities, and the severity and timing of the disturbance [2]. In clinical workflows, an ECG is often recorded before the potassium result becomes available, which has motivated interest in ECG-based tools that could prioritize testing or monitoring without waiting for the laboratory.

Artificial intelligence-enabled ECG (AI-ECG) refers to AI/machine-learning (ML) methods that learn or classify patterns from ECG data rather than relying on rule-based ECG criteria or conventional statistical regression. Recent advances in deep learning have enabled AI models to identify subtle ECG patterns that are not readily discernible by visual interpretation, creating opportunities for earlier detection of electrolyte abnormalities. These approaches have been developed using diverse ECG inputs: 12-lead, reduced-lead, single-lead, wearable, monitor-derived, and image-based recordings - and applied to four distinct tasks: detecting hyperkalemia, detecting hypokalemia, classifying potassium status into categories, and estimating a continuous potassium concentration [2-4].

Reported accuracy alone does not establish that such a model is ready for clinical use. A reported area under the receiver operating characteristic curve (AUROC), which summarizes discrimination across classification thresholds, or estimation error is difficult to interpret in isolation. Clinical utility depends on the potassium threshold, patient sampling, external validation (evaluation in an independent population beyond the development dataset), calibration (agreement between predicted and observed outcomes), and whether performance is maintained in individual patients under routine clinical conditions. Across the literature, these design features vary substantially, which makes it difficult to judge whether reported performance reflects transportable clinical accuracy or favorable development conditions.

Although narrative and modeling reviews have examined the conventional electrocardiographic changes associated with potassium disturbances [2], and a recent non-peer-reviewed preprint systematically reviewed AI-ECG across multiple electrolyte abnormalities while meta-analyzing potassium detection [5], our literature search did not identify any peer-reviewed systematic review specifically evaluating AI/ML-enabled ECG across the full spectrum of potassium-related applications. Existing syntheses have not jointly examined categorical potassium classification and continuous potassium estimation alongside detection, nor have they explicitly considered model lineage, i.e., relationships between successive versions of AI models and shared datasets, both of which may influence interpretation of the evidence.

This systematic review therefore addressed a single question: across the full spectrum of potassium-related tasks, how accurate and how clinically ready is AI-ECG when tested against a paired laboratory potassium measurement as the reference standard? We evaluated adult human evidence across four clinically relevant potassium tasks - hyperkalemia detection, hypokalemia detection, categorical potassium-status classification, and continuous potassium estimation. Particular attention was given to validation maturity (progression from model development through internal, external, prospective, and implementation evaluation), risk of bias, evidence independence, and the feasibility of quantitative synthesis.