Multi-hospital external validation of lead-i ECG AI model for the detection of moderate-to-severe hyperkalemia in chronic kidney disease (CKD) patients
N Razvi, K Lambeth, M Razavi, W L Lau, M W SegarAbstract
Background
Moderate-to-severe hyperkalemia (≥6.5 mmol/L) can be a life-threatening electrolyte disorder with mortality risk, particularly in patients with CKD. Guideline therapy for CKD can also exacerbate the condition. When detected, hyperkalemia often leads to cessation or reduction of therapy, which increases all-cause mortality risk. Diagnosis is currently confined to laboratory testing. Detection via lead-I ECG (commonly available on patient wearables) could enable remote monitoring to improve management. Validation of lead-I ECG based AI models for hyperkalemia remains limited in populations enriched for CKD across multiple health systems.
Purpose
To evaluate the performance of an AI-enabled lead-I ECG algorithm for detection of moderate-to-severe hyperkalemia (≥6.5 mmol/L) in patients with CKD from three independent health systems.
Methods
We conducted a retrospective external validation using de-identified health record data collected between 2016 and 2025 from three distinct U.S. health systems (Sanford Health, South Dakota, Sharp HealthCare, California, and Texas Health Resources, Texas). Adults aged ≥22 years with a resting 12-lead ECG and a serum potassium measurement obtained within ≤4 hours of ECG recording were included. Patients were excluded if the following were present: cardiac pacing, ventricular rate ≥120 BPM, ECG artifact, noisy signal, prolonged flat signal (e.g. due to lead disconnection). The primary endpoint was serum potassium ≥6.5 mmol/L. Performance metrics were calculated using 95% confidence intervals via Wilson method. Analyses focused on patients with estimated glomerular filtration rate (eGFR) ≤60 mL/min/1.73m².
Results
The validation cohort included 425,263 ECG–potassium pairs with eGFR ≤60 mL/min/1.73m². For detection of potassium ≥6.5 mmol/L, sensitivity was 80.53% (95% CI 0.7860–0.8234), specificity was 78.66% (95% CI 0.7854–0.7879), the area under the receiver operating characteristic curve (AUROC) was 0.8694, and the precision–recall area under the curve was 0.0520. Median prediction scores increased stepwise across potassium strata (Figure 1B), demonstrating separation between not-hyperkalemia <5.7 mmol/L (419,017 pairs, median score 0.038), mild hyperkalemia 5.7-6.49 mmol/L (4525 pairs, median score 0.813), moderate hyperkalemia 6.5-7.49 mmol/L (1319 pairs, median score 0.980), and severe hyperkalemia ≥7.5 mmol/L (402 pairs, median score 0.997). The receiver operating characteristic curve (Figure 1C) demonstrated robust discrimination in the presence of CKD and associated comorbidities.
Conclusion
In a large multi-health system geographically dispersed cohort enriched for CKD, AI-based lead-I ECG algorithm accurately detected moderate-to-severe hyperkalemia (≥6.5 mmol/L). These findings support the potential role of automated ECG analysis for non-invasive hyperkalemia screening in CKD populations.Figure 1