AI-Enabled ECG Stratification and Timing of Potassium-Lowering Therapy in Patients With Hyperkalemia: A Target Trial Emulation Study

Hyperkalaemia

artificial intelligence, electrocardiogram, emergency department, hyperkalemia, potassium, treatment

Hyperkalemia is a common and potentially life-threatening electrolyte disorder, yet there is limited evidence guiding the optimal timing of potassium-lowering therapy in routine clinical practice. Although electrocardiographic (ECG) abnormalities are recommended to inform treatment decisions, such findings are often subtle and difficult to recognize consistently by clinicians. This study aims to emulate a target trial to evaluate the association between the timing of potassium-lowering therapy (timely versus delayed initiation) and short-term mortality among patients with laboratory-confirmed hyperkalemia presenting to the emergency department. In addition, the study examines whether artificial intelligence-enabled ECG (AI-ECG) stratification identifies patient subgroups that may differentially benefit from earlier treatment. Using observational electronic health record data from multiple healthcare systems, including publicly available critical care databases and institutionally governed hospital datasets, treatment strategies are compared using causal inference methods designed to approximate randomized assignment. The primary outcome is 90-day all-cause mortality. The results of this study are intended to inform clinical decision-making regarding treatment timing in hyperkalemia and to evaluate the potential role of AI-ECG as a risk stratification tool in real-world settings.

Hyperkalemia is a frequent and clinically important electrolyte abnormality associated with increased risk of arrhythmia and mortality. Current clinical guidelines recommend considering electrocardiographic (ECG) abnormalities when determining the urgency of potassium-lowering therapy; however, specific ECG criteria prompting treatment are not uniformly defined, and subtle ECG changes are often difficult for clinicians to recognize consistently in real-world practice. As a result, substantial variability exists in the timing of treatment initiation for hyperkalemia, and randomized trials directly comparing timely versus delayed therapy are ethically and practically challenging to conduct. This study is designed as an observational study emulating a target trial to evaluate the association between treatment timing and short-term mortality among patients with laboratory-confirmed hyperkalemia. The target trial protocol was specified a priori and submitted to the Institutional Review Board of Tri-Service General Hospital and explicitly defined all major design components required to address the causal question of interest. The protocol explicitly defined all major design components prior to analysis, including the data sources, eligibility criteria, treatment strategies, assignment procedures, time zero, follow-up, outcomes, causal contrasts, and analytic methods. The emulation was conducted using routinely collected electronic health record data from three sources: Tri-Service General Hospital (TSGH), a tertiary academic medical center in Taiwan with longitudinal emergency department and inpatient records spanning January 2011 to December 2024; the Military Hospital Research Database (MHRD), which aggregates patient-level data from multiple regional military-affiliated hospitals across Taiwan during calendar year 2023; and the publicly available MIMIC-IV database, including linked emergency department, intensive care, and electrocardiogram datasets. Data from these sources were extracted, harmonized, and analyzed using a common protocol. Eligible participants were adults aged 18 years or older who presented to the emergency department with laboratory-confirmed hyperkalemia, defined by a serum potassium concentration at or above the prespecified threshold, and who underwent a standard 12-lead electrocardiogram within a defined time window around the index potassium measurement. Encounters in which potassium-lowering therapy was initiated prior to the index measurement, in which no therapy was administered within the allowable exposure window, or in which key baseline covariates required for emulating treatment assignment were unavailable were excluded. The protocol specified two treatment strategies: timely initiation of potassium-lowering therapy within a prespecified early time window following the index potassium measurement, and delayed initiation of therapy after this early window but within a maximum allowable delay period. In the hypothetical target trial, eligible patients would be randomly assigned to one of these treatment strategies at baseline. In the emulated trial, treatment assignment followed routine clinical practice and was approximated using propensity score-based methods to balance measured baseline characteristics between groups and emulate randomization. The pre-specified variables in propensity score included serum potassium (continuous), AI-ECG prediction (positive vs. negative), age (years), gender, and Charlson Comorbidity Index (CCI). A common time zero was defined as the time of the index potassium measurement. Follow-up began at time zero and continued for a fixed duration of 90 days or until death, whichever occurred first. The primary outcome was all-cause mortality within 90 days of the index event, ascertained through linkage to national death registries or in-database mortality records, depending on the data source. The primary causal contrast of interest was defined on an intention-to-treat basis, comparing the counterfactual outcomes that would have been observed had all eligible patients initiated potassium-lowering therapy in a delayed manner versus had all eligible patients initiated therapy in a timely manner. The protocol further prespecified assessment of effect modification by artificial intelligence-enabled electrocardiogram (AI-ECG) based risk stratification, with subgroup contrasts defined according to AI-ECG prediction status. Analytic methods were prespecified to estimate these causal contrasts using time-to-event analyses appropriate for observational data, following emulation of random assignment. Additional sensitivity analyses were defined to assess the robustness of findings to key assumptions and potential sources of bias, including analyses designed to mitigate immortal time bias arising from limitations in the temporal resolution of mortality data.

No related papers found yet.