Clinical decision support systems (CDSS) embedded in electronic health records (EHRs) enhance patient safety and guide clinicians toward evidence-based care¹. By leveraging “big data” and predictive algorithms, CDSS optimize workflows, reduce medication errors, and identify patients at high risk of adverse outcomes, such as in-hospital mortality^2,3. Advances in artificial intelligence (AI) have further expanded the capabilities of CDSS, resulting in innovative tools and commercial products that improve decision-making and patient outcomes. Many of these systems are now integrated into medical devices, offering new ways to visualize and interpret clinical data^4,5. However, while AI-driven CDSS hold promise, there is limited evidence supporting their economic value and clinical impact, particularly in real-world settings⁶.

The electrocardiogram (ECG) is a widely used diagnostic tool, performed over 3 million times daily worldwide⁷, and is routinely utilized for admission and preoperative evaluations to assess potential cardiac risks. With the advancement of AI⁸, AI-enabled ECG (AI-ECG) has demonstrated the ability to detect previously undetectable conditions, such as low ejection fraction and paroxysmal atrial fibrillation^9–11. AI-ECG has also shown strong predictive power for 1-year mortality, with area under the curve (AUC) values exceeding 0.85^12,13. An AUC of 0.85 indicates excellent discrimination, meaning the AI-ECG model has a high ability to differentiate between patients who will experience an event (such as death) and those who will not, highlighting its potential for use in critical care triage.

The rapid response system (RRS) has been shown to enhance the recognition and management of high-risk patients, reducing mortality^14,15. A recent randomized controlled trial (RCT) evaluated the integration of AI-ECG into a track-and-trigger system (TTS) within a RRS. This innovative approach enabled real-time identification of hospitalized patients at high mortality risk, prompting physicians to initiate timely interventions¹⁶. The trial demonstrated a significant 17% reduction in 90-day mortality (hazard ratio [HR]: 0.83; 95% confidence interval [CI]: 0.70–0.99) in the intervention group compared to controls. However, this improvement was accompanied by increased ICU admissions and additional cardiac investigations, leading to higher care cascade costs.

Despite these promising results, the cost-effectiveness of AI-ECG systems remains unclear, particularly in the context of their integration into routine clinical practice. Given that healthcare resources are limited, investments in new technologies often come at the expense of potential investments in other areas, highlighting the importance of considering opportunity costs. Therefore, it is crucial to evaluate not only clinical efficacy but also the economic impact of new technologies to support informed healthcare decision-making. This study is novel in providing the first comprehensive economic evaluation of an AI-ECG system, assessing both its clinical outcomes and incremental cost-effectiveness from a health payer’s perspective. By addressing these critical gaps, this research aims to inform the sustainable adoption of AI-driven CDSS in real-world healthcare settings.

The present study reports the direct medical costs associated with implementing AI-ECG alert system in predicting mortality risk compared to usual care in hospitalized patients from a RCT. Results of this economic evaluation revealed the implementation of AI-ECG resulted in a modest increase in short-term healthcare costs while reducing 90-day all-cause mortality. From a healthcare payer’s perspective, the intervention appears to represent a cost-effective strategy, particularly when used in a population with elevated baseline risk.

Using AI to analyze existing ECGs can provide clinicians with a wealth of additional information. Beyond indicating high mortality risk^12,13, AI can also identify conditions like low ejection fraction and paroxysmal atrial fibrillation^9–11, prompting clinicians to consider further interventions. Previous RCT have demonstrated that this type of AI-ECG can prompt clinicians to arrange echocardiograms for high-risk patients to confirm the possibility of low ejection fraction¹⁷. The increase in interventions following such alerts is not unique to AI-ECG; other studies have also shown that AI alerts can influence physician decision-making¹⁸. The present findings also indicate that the AI-ECG alert system for mortality risk may alter physician behavior, such as by prompting more diagnostic tests or increasing the provision of intensive care¹⁶. In the present study, the observed differences in total mean costs per patient were primarily associated with higher drug costs, followed by ICU stay costs and medical material costs. Therefore, the focus of the cost-effectiveness analysis is on the relationship between the costs of these additional medical interventions and the observed reduction in mortality.

In the base-case analysis, the ICER was $59,500 per death averted. The cost-effectiveness acceptability curve indicated a 95% probability that the intervention would be considered cost-effective at a WTP threshold of $409,321. This wide range of uncertainty is likely due to the relatively small differences in both costs and mortality between the intervention and control groups. When such differences are modest, even small variations in either can cause large fluctuations in the ICER estimates, especially when using resampling methods. The base-case ICER corresponds to approximately 1.8 times Taiwan’s WTP threshold for a QALY gained, while the upper bound aligns with 12.3 times that threshold. Given the average age of patients in the trial (61 years) and a life expectancy of around 80 years in Taiwan, each death averted could conservatively translate into 7 to 12 additional QALYs. Using a WTP threshold equivalent to one time Taiwan’s GDP per capita, the implied acceptable range for a death averted would be between $232,638 and $398,808. In this context, the findings suggest that the intervention is likely to be cost-effective in Taiwan’s healthcare system and potentially applicable to other middle- or high-income countries with similar or higher WTP thresholds.

In this cost-effective analysis, the cost of the AI-ECG alert system itself is the first factor to consider. However, the results of this study suggest that the costs of the AI-ECG alert system have no significant impact on healthcare team expenditures. Since ECGs are inherently very inexpensive tests, past research has emphasized that the cost-effectiveness analysis of ECGs should focus more on the care cascades triggered by ECGs. In asymptomatic adults, low-value ECGs can initiate care cascades that lead to substantial economic burdens^19,20. However, the in-hospital population targeted in this study had a higher mortality rate (4.3% in the control group), making ECG both reasonable and in line with standard clinical criteria²¹. In the base-case analysis, each AI-ECG alert was priced at $4, equivalent to the cost of a standard ECG, and gradually increased up to eightfold. Despite this, the ICERs remained stable, suggesting that the system’s cost-effectiveness is not significantly impacted by the cost associated with AI analysis itself.

Many clinical support systems have been designed to assist physicians and improve patient safety, yet these alerts are often overridden due to alert fatigue. Previous studies have shown that override rates can reach up to three-quarters in inpatient settings and nearly two-thirds in emergency departments^22,23, though most of these were related to prescribing alerts. As most costs stem from the testing and treatment following an AI-ECG alert, even if a higher override rate at 60% is seen with large-scale implementation in the future, while the mortality reduction effect may weaken, the ICER would remain relatively stable, indicating that the economic benefits of implementing AI-ECG alerts are unlikely to be diminished by alert fatigue, in part because the additional costs of the AI-enabled ECG system is relatively low.

Critical care, particularly admission to ICU, is widely recognized for significantly reducing patient mortality²⁴. An important issue is the extremely high cost of ICU care, and numerous studies have previously focused on identifying patients who would benefit most cost-effectively from ICU interventions. Cost-effectiveness analysis has shown that ICU intervention is economically beneficial for patients with a Simplified Acute Physiology Score (SAPS) II predicted mortality greater than 5%, and it is even more cost-effective for those with predicted mortality exceeding 40%²⁵. Another study suggests considering the withdrawal of ICU care for patients with an Acute Physiology and Chronic Health Evaluation (APACHE) III predicted mortality greater than 90%²⁶. Two seemingly opposing pieces of evidence highlight a key point: interventions should be directed toward patients with a moderate risk of mortality. In the control group of this study, the high-risk patients had a mortality rate of 23%, while low-risk patients had a 2.4% mortality rate, aligning with the recommendations of previous studies²⁷. A study has reported that 62% of inpatients who experienced cardiac arrest showed signs of physiologic instability for more than 6 h before the event, yet only 22% had explicit documentation²⁸. This indicates that critical care should focus on the early identification of deterioration signs, rather than recognizing them only when the patient enters an irreversible process of dying²⁹. By appropriately allocating medical resources to patients showing early signs of deterioration, mortality rates can be reduced. This study demonstrates that AI-ECG can substantially enhance the cost-effectiveness of critical care.

A key limitation of the study is that it was conducted in Taiwan, with cost analysis tailored to the Taiwanese healthcare system, which may restrict the generalizability of the findings to other countries. Differences in healthcare systems, along with demographic and socioeconomic disparities, could influence the cost-effectiveness. Previous studies comparing AI-ECG for asymptomatic left ventricular dysfunction screening to usual care have shown significant variations in ICERs, primarily due to differences in healthcare costs across regions^30,31. Therefore, further investigation and validation across diverse healthcare settings are necessary to ensure broader applicability. Moreover, although development, staff training, maintenance, and administrative costs are important components of a comprehensive cost-effectiveness analysis, these implementation-related expenses were not included in the base-case analysis. Omitting such costs may lead to an underestimation of the ICER, as they contribute meaningfully to the real-world cost and effectiveness of the intervention³². However, these expenses are often difficult to quantify during early-stage development and are likely to become clearer during commercialization and reimbursement planning. Importantly, our sensitivity analyses showed that the ICER per death averted remained stable even when the cost of the AI-ECG system was increased by up to eightfold, suggesting that the findings are robust to variations in the cost of the intervention itself.

Additionally, the evaluation was conducted from a health payer’s perspective, excluding costs borne by patients and their families after hospitalization, which may result in an underestimation of the total economic impact. We also did not include future healthcare costs associated with extended survival (i.e., life extension costs), which are sometimes considered in lifetime models. This exclusion may further contribute to a conservative estimate of the ICER. Lastly, this study did not extrapolate deaths averted into QALYs gained or adopted a lifetime horizon, as the trial population was heterogeneous, and the follow-up period was limited to 90 days. Future research may consider these approaches to capture longer-term health outcomes.

The AI-ECG alert significantly reduced all-cause mortality among hospitalized patients. Although it incurred slightly higher costs than usual care, the incremental costs per death averted suggest favorable cost-effectiveness, even when accounting for the large uncertainty in the ICER. The AI-ECG alert system presents a promising and cost-effective approach to improving patient outcomes. However, further validation across diverse clinical settings is necessary to ensure the generalizability of these findings.

The economic evaluation was a cost-effectiveness analysis conducted alongside a pragmatic randomized clinical trial in Taiwan, assessing the impact of an AI-ECG alert system on clinical outcomes¹⁶. The study period ran from December 15, 2021, to April 30, 2022, with follow-up extending to July 31, 2022, allowing for a 90-day post-intervention analysis for all patients included up to the end of the intervention period. The Consolidated Health Economic Evaluation Reporting Standards (CHEERS) checklist was used to demonstrate adherence to its reporting guidelines³³.

This study was supported by funding from the National Science and Technology Council, Taiwan (NSTC 113-2320-B-016-011 to P.H.H.). The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

P.H.H., C.L., and C.S.T. conceptualized the study. P.H.H., C.L., T.K.L., and C.S.T. designed the methodology. C.S.L. and W.T.L. did the data collection. P.H.H., C.L., and D.J.T. did the data analysis. P.H.H. wrote the original draft of the manuscript. All authors reviewed and edited the manuscript. C.S.T. supervised the data collection and analyses. P.H.H., C.L., W.T.L., and C.S.T. did project administration. Y.J.H., Y.H.C., C.Y.L., S.H.L., and C.S.T. acquired funding. P.H.H. and C.L. verified the data. All authors have read and agreed to the final version of the manuscript and to the submission for publication.

Patient data cannot be made publicly available due to privacy concerns. De-identified tabular data can be obtained from the corresponding author on approval from the ethics committee of the Tri-Service General Hospital. Approval from this committee can be requested from the Tri-Service General Hospital’s Clinical Trial Management System (https://tsgh.cims.tw/wiPtms/index.html), with an expected review period of approximately 2–3 months. After approval, researchers will be granted VPN access to perform analyses, ensuring data security and confidentiality (summary data can be exported), with measures in place to prevent any breach of personal information.

The authors declare no competing interests.

The online version contains supplementary material available at 10.1038/s41746-025-01735-7.