AI-ECG mortality alert: a 15,965-patient randomized trial in Taiwan cut 90-day all-cause mortality from 4.3% to 3.6% (HR 0.83)

Verification status: IN CHECKING — not publish-ready, not pending, not verified. This rests on a single peer-reviewed randomized trial run at one hospital system in Taiwan. Peer review is not independent re-measurement, and no external site has yet replicated the effect. Read the caveats below.

The problem

Identifying which hospitalized patients are about to deteriorate is, in the authors’ words, a task that “poses a substantial challenge in clinical practice” (source). A standard 12-lead ECG is cheap, fast and taken routinely, but the mortality signal buried in it is not something a busy physician reliably reads off the tracing in time to act.

What was built

The team built an AI-enabled ECG system that scores a patient’s mortality risk from the ECG and, for patients it flags as high risk, surfaces “an AI report and warning messages delivered to the physicians, flagging patients predicted to be at high risk of mortality” (source). Crucially, the alert is a decision-support nudge to a human, not an autonomous action: the physician still decides what care to arrange. The evaluation was “a multisite randomized controlled trial involving 39 physicians and 15,965 patients” (source), registered as ClinicalTrials.gov NCT05118035. The independent outlet News-Medical restated the arm sizes: “The intervention group included 8,001 patients, while the control group had 7,964 patients” (source).

The outcome

The trial met its primary outcome. Per the paper’s abstract, “implementation of the AI-ECG alert was associated with a significant reduction in all-cause mortality within 90 days: 3.6% patients in the intervention group died within 90 days, compared to 4.3% in the control group (4.3%) (hazard ratio (HR) = 0.83, 95% confidence interval (CI) = 0.70-0.99)” (source). The independent outlet HealthManagement.org restated the same headline figures: “3.6% of patients in the intervention group succumbing within the specified timeframe, compared to 4.3% in the control group” (source).

The benefit was not spread evenly; it was concentrated where the model raised its hand. A prespecified analysis found “reduction in all-cause mortality associated with the AI-ECG alert was observed primarily in patients with high-risk ECGs (HR = 0.69, 95% CI = 0.53-0.90)” (source), an effect HealthManagement.org summarized as a “31% reduction in deaths” (source). A second independent outlet, HCPLive, restated the same subgroup figure: “the reduction in all-cause mortality associated with the AI-ECG alert was identified primarily among those with high-risk ECGs (HR, 0.69; 95% CI, 0.53 - 0.90)” (source). Within that high-risk group, the sharpest movement was in cardiac death: the paper reports “a significant reduction in the risk of cardiac death (0.2% in the intervention arm versus 2.4% in the control arm, HR = 0.07, 95% CI = 0.01-0.56)” (source). That last confidence interval is very wide (0.01-0.56), reflecting the small number of cardiac deaths, so the point estimate should be read as directional rather than precise.

Weakest load-bearing source, named. The exact figures come from a single trial run at one hospital system, Tri-Service General Hospital / National Defense Medical Center in Taiwan, and were verified here against the MEDLINE-indexed abstract on PubMed, not the paywalled Nature Medicine full text (source). The three independent outlets that restate the numbers are reporting on that same trial, not re-measuring it, so their agreement confirms the figures were reported accurately, not that the effect reproduces elsewhere. No party unaffiliated with the trial has yet replicated the result. This is a strong primary corroborated by independent reporting, not an independently validated multi-site finding.

How this was verified. Method: the 3.6%-vs-4.3% primary result, the 39-physician / 15,965-patient counts, the high-risk subgroup HR (0.69), and the cardiac-death figures were quote-matched verbatim against the MEDLINE-indexed abstract of the Nature Medicine paper (published 29 April 2024, DOI 10.1038/s41591-024-02961-4; PMID 38684860; NCBI eutils capture saved to sources/pubmed-38684860-abstract.txt; Wayback 20260828064453 — Tier 1, peer-reviewed pragmatic RCT). The headline mortality figures and the arm sizes are independently restated by HealthManagement.org (Wayback 20260828064544) and News-Medical (Wayback 20260828064512), and the high-risk subgroup HR 0.69 is restated verbatim by a third independent outlet, HCPLive (Wayback 20250907070212), all Tier 2. No confirmation was sought from the authors or the hospital: asking the subject to confirm its own figures is a testimonial, not an audit.

Sources

1. Nature Medicine · AI-enabled electrocardiography alert intervention and all-cause mortality: a pragmatic randomized clinical trial · Lin CS, Liu WT, Tsai DJ, et al. · 29 April 2024 · https://pubmed.ncbi.nlm.nih.gov/38684860/ — Tier 1 (primary, peer-reviewed pragmatic RCT; DOI 10.1038/s41591-024-02961-4, Nat Med 2024;30(5):1461-1470; ClinicalTrials.gov NCT05118035; source of the 3.6%/4.3% primary result, HR 0.83, the high-risk HR 0.69 and the cardiac-death figures; verified via the MEDLINE-indexed abstract, NCBI eutils capture in sources/pubmed-38684860-abstract.txt, archived Wayback 20260828064453).
2. HealthManagement.org · AI-Enabled ECG Reduces Mortality: Breakthrough in Medical AI · 30 April 2024 · https://healthmanagement.org/c/it/news/ai-enabled-ecg-reduces-mortality-breakthrough-in-medical-ai — Tier 2 (independent health-technology outlet; independently restates the 3.6% vs 4.3% mortality figures, the 39-physician/15,965-patient counts and the ~31% high-risk reduction; archived Wayback 20260828064544, saved to sources/healthmanagement-20240430.html).
3. News-Medical.net · AI-enabled ECG system significantly reduces hospital mortality rates by identifying at-risk patients · 30 April 2024 · https://www.news-medical.net/news/20240430/AI-enabled-ECG-system-significantly-reduces-hospital-mortality-rates-by-identifying-at-risk-patients.aspx — Tier 2 (independent science-news outlet; independently restates the arm sizes 8,001 intervention / 7,964 control and the multisite Taiwan setting; archived Wayback 20260828064512, saved to sources/news-medical-20240430.html).
4. HCPLive (MJH Life Sciences) · AI-Enabled Electrocardiogram Alert Intervention Reduces All-Cause Mortality · Connor Iapoce · 1 May 2024 · https://www.hcplive.com/view/ai-electrocardiogram-alert-intervention-reduces-all-cause-mortality — Tier 2 (independent physician-facing medical outlet; independently restates the primary 3.6% vs 4.3% mortality result and the high-risk-ECG subgroup HR 0.69, 95% CI 0.53-0.90 verbatim, attributed to “Lin and colleagues”; archived Wayback 20250907070212, saved to sources/hcplive-20240501.html).

Related case files

• Bayesian Health / Johns Hopkins TREWS, a sepsis AI that also moved a hard mortality endpoint — the closest comparator: another physician-facing AI alert that reduced measured mortality, not just a process metric.
• The Epic Sepsis Model external-validation failure — the cautionary contrast, and exactly why the “path to green” here demands external replication: a widely deployed clinical AI that performed far worse at an independent site than where it was built.
• UC San Diego’s LLM sepsis-feedback RCT, where process improved but mortality did not — the mirror image: a randomized clinical-AI trial that lifted a compliance metric yet showed no mortality benefit, unlike this one.

Path to green

The independent public record would have to yield an external replication of the AI-ECG mortality effect at a hospital system unaffiliated with National Defense Medical Center / Tri-Service General Hospital, or an independent re-analysis of the trial data by investigators with no authorship on the primary paper. Either would raise this above “single-system primary plus independent reporting.” A registry or health-system audit of before-and-after mortality following deployment, reported by a party with no stake in the tool, would do the same. No contact should be chased; absent those independent documents the story caps at its achieved band.