AI ECG Potassium Alerts: What the ED Trial Showed

A 72-year-old retired welder on hemodialysis comes in at 0200, weak and short of breath, after missing his Saturday session. The triage ECG crosses at 0207. The T waves look a little tall in V2 and V3, but no taller than they looked last month, and the QRS is narrow. The potassium will not result for at least 40 minutes. In one Taiwanese ED system, a red pop-up would already be on the physician's screen reporting an ECG-estimated potassium and a high-risk flag.

Bottom line: the only randomized trial of an AI-ECG potassium alert in emergency physicians shows that it speeds hyperkalemia treatment in the small group it flags. Use it as a hyperkalemia prompt tied to a written treat-and-confirm pathway and a local accuracy audit, never as a replacement for the potassium level or as a hypokalemia tool.

What the trial tested

Lin and colleagues ran a pragmatic, open-label randomized trial in the emergency departments of two Tri-Service General Hospital system hospitals in Taiwan, one academic and one community [1]. Randomization was by physician: 70 emergency physicians, 35 per arm, from January to October 2022, with a pause from May through August for COVID-19 [1]. Adults with at least one ECG were included, and only the first ECG counted. Patients under 18 and patients treated for a potassium disorder before their ECG were excluded, which left 7,506 patients in the alert arm and 7,483 in the control arm [1].

Every ECG ran through the model. Intervention physicians received a scarlet pop-up in the electronic record when the model predicted potassium of 6.0 mmol/L or higher or 3.0 mmol/L or lower, along with the estimated value [1]. The investigators set the alert thresholds to produce a positive predictive value of at least 40% to limit alert fatigue [1]. Control physicians saw nothing, and their patients' predictions were stored for later analysis [1].

The two primary outcomes were hyperkalemia-related treatment and hypokalemia-related treatment within 3 hours. Hyperkalemia treatment counted IV calcium chloride, insulin with glucose, sodium bicarbonate, loop diuretics, inhaled beta-agonists, oral exchange resin, and hemodialysis [1]. The key safety outcome was overcorrection within 24 hours: potassium of 3.5 mmol/L or lower after hyperkalemia treatment, or 5.5 mmol/L or higher after supplementation [1].

Where the alert changed care

Across all patients, the alert changed nothing measurable. Hyperkalemia-related treatment was 8.0% vs 7.7% (HR 1.05, 95% CI 0.94 to 1.17) and hypokalemia-related treatment was 2.1% vs 2.4% (HR 0.91, 95% CI 0.74 to 1.13) [1]. The authors expected a diluted overall result and prespecified an analysis of the patients the model flagged [1].

The flagged hyperkalemia group was small: 81 patients in the alert arm and 77 in the control arm, about 1% of each [1]. Inside that group, treatment within 3 hours was 69.1% vs 41.6% (HR 2.23, 95% CI 1.44 to 3.46), and treatment within 1 hour was 42.0% vs 15.6% (HR 3.13) [1]. Among patients not flagged for hyperkalemia, there was no difference [1].

The effect was largest where human ECG reading is weakest. When the model estimated moderate hyperkalemia, treatment was 62.1% vs 33.3%. When it estimated severe hyperkalemia (6.5 mmol/L or higher), treatment was 82.6% vs 62.5% [1]. Most of the difference came before the lab resulted: 38.3% vs 18.2% treated before the potassium report, compared with 30.9% vs 23.4% after it [1]. Alerted physicians also ordered more potassium tests, which is the behavior you want [1].

The hypokalemia alert did nothing. In the flagged hypokalemia group, treatment within 3 hours was 31.4% vs 29.7% (HR 1.14, 95% CI 0.77 to 1.77) [1].

What the trial did not show

Positive predictive value in practice was 43.6% for the hyperkalemia flag and 42.7% for the hypokalemia flag [1]. The authors note that prevalence was likely underestimated, because many flagged patients were treated before blood was drawn: only 55 of 158 flagged hyperkalemia patients had a paired level within an hour [1]. On the measured numbers, more than half of alerts were false positives at that operating point.

Overcorrection was rare and did not differ between arms: 27 vs 32 treatment-induced hypokalemia events overall, and 2 vs 1 in the flagged hyperkalemia group [1]. Both alert-arm cases followed emergent dialysis for potassium above 7.0 mmol/L [1].

No clinical outcome improved. Cardiac arrest was 49 vs 44 (HR 1.11, 95% CI 0.74 to 1.67). All-cause death within 365 days was 692 vs 631 (HR 1.11, 95% CI 1.00 to 1.24, p = 0.053), slightly higher in the alert arm [1]. The trial was powered for treatment rates, not death, so that number answers no question. The specific uncertainty is that nobody has shown this alert helps patients survive.

Generalizability is the other gap. Most patients came from a single academic center, the model was trained on that same institution's data, and the investigators did not log whether physicians acknowledged each alert [1]. The same group's earlier trial of a mortality-risk AI-ECG alert in 15,965 hospitalized patients did reduce 90-day death (3.6% vs 4.3%, HR 0.83, 95% CI 0.70 to 0.99), but that was a different model with a different target [2].

Why the plain ECG needs help

The ECG alone misses most hyperkalemia. In a retrospective review of 90 inpatients with potassium of 6.0 or higher and a concurrent ECG, 52% had any ECG change and only 18% met strict criteria for new peaked, symmetric T waves that resolved after correction [3]. The authors found no support for using ECG changes to guide treatment of stable patients [3].

AI models do better, but their performance varies by site. A Mayo Clinic model built on two leads, validated in 61,965 patients with stage 3 or worse chronic kidney disease across three Mayo sites, had an AUC of 0.853 to 0.883 [4]. At a 90% sensitivity operating point, specificity ranged from 54.7% in Florida to 63.2% in Minnesota [4]. The same model, at the same threshold, produced different false-positive burdens in different hospitals.

Build the pathway before you switch it on

Write the response first, using criteria from the 2023 UK Kidney Association guideline [5]:

• On an alert: put the patient on continuous cardiac monitoring, get the 12-lead in front of a physician, and send a potassium level with a stated turnaround target. The guideline recommends continuous monitoring for potassium of 6.5 mmol/L or higher, for any hyperkalemic ECG change, and for 6.0 to 6.4 mmol/L in an unwell patient or when a rapid rise is expected [5].
• ECG changes present: give IV calcium, 30 mL of 10% calcium gluconate over 10 minutes or 10 mL of 10% calcium chloride over 5 minutes. Chloride is preferred in arrest or peri-arrest, and the guideline does not recommend calcium with a normal ECG [5].
• Insulin-glucose: 10 units of soluble insulin with 25 g of glucose IV, recommended for potassium of 6.5 or higher and suggested for 6.0 to 6.4 [5]. Check glucose first. If it is below 7.0 mmol/L (about 126 mg/dL), follow with 10% glucose at 50 mL per hour for 5 hours, and check glucose at 0, 30, 60, 90, 120, 180, 240, 300, and 360 minutes [5]. Renal failure, low body weight, older age, and no diabetes all raise hypoglycemia risk [5].
• Recheck potassium at 1, 2, 4, 6, and 24 hours after treatment for moderate or severe hyperkalemia [5].

Then audit locally. Before go-live, run the model silently on your own ECGs paired with your lab values. After go-live, review flagged patients monthly and report positive predictive value, time to treatment, hypoglycemia, and overcorrection.

Key Takeaways

• The only emergency physician RCT of an AI-ECG potassium alert raised hyperkalemia treatment within 1 hour from 15.6% to 42.0% in flagged patients, with no effect in the overall population.
• Expect more than half of hyperkalemia alerts to be false positives: the trial's positive predictive value was 43.6%.
• Use the alert as a hyperkalemia prompt tied to a written treat-and-confirm pathway and a local accuracy audit, and do not deploy it for hypokalemia.
• Before insulin-glucose, check glucose. If it is below 7.0 mmol/L, run 10% glucose at 50 mL per hour for 5 hours, and monitor glucose for 6 hours.

FAQ

Can AI detect hyperkalemia from an ECG?

Yes, with error. In a 2022 emergency department trial, the model's AUC for potassium of 6.0 mmol/L or higher was 0.971, but only 43.6% of hyperkalemia alerts were confirmed by a paired lab value [1]. Treat the alert as a reason to get a level and look at the ECG again.

Do AI-ECG potassium alerts improve patient outcomes?

Not yet shown. The randomized trial found faster treatment in flagged patients but no difference in cardiac arrest, and 365-day death was 692 vs 631 (HR 1.11, 95% CI 1.00 to 1.24) [1]. The trial was not powered for outcomes.

Should you treat hyperkalemia before the potassium result comes back?

Give IV calcium when the ECG shows hyperkalemic changes, including isolated peaked T waves, without waiting for the lab [5]. The 2023 UK Kidney Association guideline ties insulin-glucose to a measured potassium of 6.0 mmol/L or higher [5]. An AI estimate alone does not meet that threshold.

Dr. Chet's Take

I have been an ED attending for more than 25 years, and the gap between the ECG and the potassium result is as old as the specialty. As a TelEmergency medical director, I also answer for rural departments where one physician reads that ECG with nobody to ask. This trial gets the important part right. The authors tested what physicians did with the output, prespecified the flagged group, and counted actual orders instead of model accuracy. The core argument holds. In the patients it flags, the alert moves calcium and insulin earlier, and a measured change in physician behavior from an AI tool is still rare. Most vendors show me an AUC and call it evidence.

That being said, this is a process result from one health system, with a home-grown model and physicians who knew they were being watched. The honest answer is that I do not know whether a pop-up that speeds treatment saves anyone, and neither does anyone else yet. I worry more about month six than month one. A tool that is wrong more often than it is right trains a department to click past it, and the dismissed alert will eventually belong to the dialysis patient who looks fine. The model is the easy part to buy. The response to every false alarm, at 0300, with a full waiting room, is the part you have to build and keep honest. That work belongs to nursing, pharmacy, and the lab as much as to the physician.

If you are the medical director being asked to turn this on, ask for three things first. Get the false-positive rate measured on your own ECGs, paired with your own lab values, in your own patients. Name one physician who reviews every flagged chart for the first 90 days. Build an order set that starts with a monitor, a stat level, and calcium when the ECG shows changes, and put the glucose checks in that same order set so nobody has to remember them on a busy night. Tell the nurses what the alert means before it fires. If the vendor cannot produce the first item, you would be running their validation study on your patients. Do not switch it on until all three exist.

— Chester Shermer, MD, FACEP | Emergency Medicine, 25+ Years Clinical Experience | State Surgeon

Continue Your Training

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download the AI in EM Survival Guide here.

Browse All Courses at Global MedOps Command

Relevant Reading on Global MedOps Command:

How to Avoid Becoming an AI Casualty — Dr. Shermer's guide to navigating AI tools in clinical and operational settings without compromising judgment or patient outcomes.

Emergency Department Efficiency Playbook — Practical systems for throughput, triage optimization, and operational efficiency.

Read more from Dr. Shermer on Medium

Connect with Dr. Shermer: LinkedIn — Chester "Chet" Shermer, MD, FACEP.

References

1. Lin C, Lin CS, Chen SJ, et al. AI-enabled electrocardiogram alert for potassium imbalance treatment: a pragmatic randomized controlled trial. Nat Commun. 2026;17(1):159. PubMed 41507124
2. Lin CS, Liu WT, Tsai DJ, et al. AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial. Nat Med. 2024;30(5):1461-1470. PubMed 38684860
3. Montague BT, Ouellette JR, Buller GK. Retrospective review of the frequency of ECG changes in hyperkalemia. Clin J Am Soc Nephrol. 2008;3(2):324-330. PubMed 18235147
4. Galloway CD, Valys AV, Shreibati JB, et al. Development and validation of a deep-learning model to screen for hyperkalemia from the electrocardiogram. JAMA Cardiol. 2019;4(5):428-436. PubMed 30942845
5. Alfonzo A, Harrison A, Baines R, et al. Clinical practice guidelines: treatment of acute hyperkalaemia in adults. UK Kidney Association; 2023. UK Kidney Association guideline

Related simulation training: EM-Sim · Books by Dr. Shermer