Taiwan AI-ECG alerts reduce mortality in a randomized trial

The AI-ECG mortality alert system at Tri-Service General Hospital. [1] [3]

A convolutional survival network, documented in ECGSurvNet. [1] [3]

The trial’s model weights are proprietary. No release identifier is published. [1] [3]

National Defense Medical Center and Tri-Service General Hospital. [1] [3]

The model estimated mortality risk from electrocardiograms and identified patients for alerts. [1] [3]

Philips PageWriter TC30 and TC50 machines recorded 10-second ECGs at 500 Hz. An NVIDIA DGX-1 in the hospital’s private cloud processed them. Alerts used the 95th percentile of risk. [1] [3]

The model analyzed 12-lead ECG waveforms and returned a risk percentile. High-risk results triggered physician alerts. [1] [3]

Physicians reviewed the ECGs and decided on further tests and treatment. The software sent alerts and did not prescribe care. [1] [3]

Routine ECGs were analyzed automatically, with warnings sent to physicians and risk reports available in the electronic health record. Physicians assessed alerted patients and selected care using their clinical judgment.

Physicians reviewed the alerts and chose further tests and treatment. The system did not prescribe or administer care. [1]

A convolutional neural network assessed ECGs and sent high-risk alerts to physicians. Clinicians selected further assessment and treatment. The trial evaluated this combined care pathway. The precise mechanism of the survival improvement remains uncertain.

A convolutional neural network assessed ECGs and sent high-risk alerts to physicians. Clinicians selected further assessment and treatment. The trial evaluated this combined care pathway. The precise mechanism of the survival improvement remains uncertain.

AIBD-0578-outcome-01 [1]

Fewer patients died within 90 days when their physicians received AI-ECG alerts. [1]

Improved Physical Health

Observed in a randomized trial. [1]

People.

Adult emergency and hospitalized patients whose physicians participated in the trial. [1]

The analysis included 8,001 patients assigned to AI alerts and 7,964 assigned to usual care. It does not identify individual lives saved. [1]

The trial included an academic center and a community hospital. Most patients came from the academic center, and the mortality difference was larger in the prespecified high-risk group. [1]

Mortality was measured for 90 days after enrollment. [1]

Yes

The alert could lead to additional tests and treatment. The economic analysis found higher mean medical costs. [1]

NCT05118035 [1]

An unblinded, pragmatic randomized trial with patients assigned to alerts or usual care. Allocation used a random list of medical-record numbers prepared before enrollment. [1]

Researchers compared deaths recorded in hospital records within 90 days. Patients were censored at their last known follow-up. [1]

The analysis included 15,965 adults treated by 39 physicians at two hospitals in Taiwan. [1]

All-cause mortality within 90 days was the primary outcome. [1]

Usual care without the AI-ECG report or high-risk alert during the 90-day follow-up. [1]

Mortality was 3.6% with alerts and 4.3% with usual care. The hazard ratio was 0.83 (95% CI, 0.70–0.99). [1]

Randomization supports an effect of the alert-assisted care pathway. It does not identify which subsequent clinical actions produced the survival difference. [1]

The study was unblinded and concentrated in one health system. Deaths outside its records could be missed. Survival beyond 90 days was not established. [1]

Patients entered from 15 December 2021 through 30 April 2022. A physician received an alert for the first high-risk ECG. [1]

The article includes a protocol and randomization details. Training code is public. Patient data require ethics approval and secure access. [1] [2] [3]

Medical expenditure and additional cost per death averted. [2]

Mean medical costs were US$6,204 with AI and US$5,803 with usual care. The base-case estimate was US$59,500 per death averted. [2]

Hospital costs were recorded in New Taiwan dollars and reported in U.S. dollars using a Bank of Taiwan rate dated 18 November 2024. The paper does not state the numerical exchange rate. [2]

Mean cost per patient over 90 days, and incremental cost per death averted. [2]

Hospital expenditure was observed. The price of the AI service and the cost-effectiveness calculations were modeled. [2]

Taiwan’s National Health Insurance was the stated perspective. Hospital records included insurance reimbursement and patients’ out-of-pocket spending. [2]

These figures describe additional expenditure for a survival benefit. They are not cash savings or net financial returns. [2]

The analysis included hospital medical costs and assumed US$4 per AI-ECG service. It excluded development, staff training, maintenance, administration, and wider family or societal costs. [2]

Patients randomly assigned to usual care supplied the comparison for costs and mortality. [2]

The authors divided additional costs by deaths averted and used 5,000 bootstrap samples to estimate uncertainty. [2]

The study did not estimate profit, a financial return, or a payback period. [2]

Researchers at the National Defense Medical Center and Tri-Service General Hospital developed the system. Physicians at an academic center and a community hospital provided care. [1]

The developers and both hospitals were in Taiwan. [1]

The reviewed clinical sources identify the organizations involved, but do not establish their complete ownership or control arrangements. [1]

The trial used Philips 12-lead ECG machines, automatic uploads, a private cloud with an NVIDIA DGX-1, electronic records, and smartphone alerts. [1]

The trial took place in Taiwan. The paper describes a private cloud, without reporting an international transfer of patient data. [1]

Not assessed

Reliable ECG transfer, access to the risk report, and physicians able to assess and treat patients were part of the tested pathway. [1]

The reviewed sources do not report an independent repetition of this clinical result. [1]

The 2025 publication analyzes costs from the same trial. It does not extend clinical follow-up beyond 90 days. [2]

The trial used a licensed model. Public architecture and training code do not include the exact weights deployed in this trial. [1]

Participating hospitals tested whether mortality alerts improved survival. [1]

The trial protocol specified random allocation and a 90-day mortality outcome. [1]

Hospitals provided alerts to the intervention group while continuing usual clinical care. [1]

The prespecified primary outcome was all-cause mortality within 90 days. [1]

3 retained sources.

Original citations and the passages that support each finding. [1] [2] [3]

English. No translation was used. [1]

Developer, Operator [1] [2]

The original studies and their supporting materials are linked to the findings that they support. [1] [2] [3]

Access and version details are listed for each reviewed source. [1] [2] [3]

AI-ECG alerts reduced 90-day mortality in the randomized comparison. [1]

The randomized trial supports the mortality comparison. The later analysis supplies the financial figures. [1] [2]

A peer-reviewed randomized trial by the developers and deploying hospitals. The cost study reuses the same trial. [1]

The precise deployed weights, longer-term survival, and the clinical actions that caused the mortality difference remain unresolved. [1]

The benefit described here, for the beneficiary groups shown.

Depth 5 for a measured survival benefit.

Depth 5 for a measured survival benefit.

The depth describes the reported result. Evidence strength is assessed separately.

AIBD’s five-level depth rubric, reviewed on 25 October 2026.

AIBD, 25 October 2026.

The follow-up supports survival over 90 days, not a lifetime estimate.

A peer-reviewed randomized trial by the developers and deploying hospitals. The outline remains appropriate under the source-independence rule.

A randomized clinical trial and an economic analysis of the same patients. [1]

Not assessed

Healthcare - mortality prevention. ECG-based clinical decision support.

Identify patients at high risk of death and prompt timely clinical assessment.

Enrollment ran from 15 December 2021 to 30 April 2022, with 90 days of follow-up. [1]

Clinical use within a trial at two hospitals in Taiwan. [1]

SHA-256: cc0075c37b4f…

NCT05118035 [1]

Not assessed

Not assessed

3 revisions. Latest: 25 October 2026. [1] [2]