Timely recognition and intervention in clinically deteriorating ward patients are essential for preventing adverse outcomes such as unexpected cardiac arrest and unplanned intensive care unit (ICU) admissions. Traditional early warning systems (EWS), including the Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS), rely on simple physiological parameters and have demonstrated utility in detecting at-risk patients. However, these approaches often suffer from limited predictive accuracy and inconsistent responsiveness in real-world practice [1,2,3,4].

To address these limitations, a variety of artificial intelligence (AI)-based EWS have been developed globally, integrating large volumes of electronic medical record (EMR) data to improve predictive performance [5,6,7,8]. The Deep learning–based Cardiac Arrest Risk Score (DeepCARS), developed in Korea, is one such system that has demonstrated high sensitivity and specificity in prior validation studies [9,10,11]. However, high predictive accuracy alone does not guarantee successful clinical implementation, as challenges such as alert fatigue and workflow integration issue remain significant barriers in real-world practice [12]. Therefore, understanding how clinicians actually interpret and respond to these alerts is a crucial, yet understudied, aspect of their clinical utility. Although primarily designed for cardiac arrest prediction, its threshold-based alerts may also influence broader physician decision-making processes, including ICU transfer, specialty consultations, or even limitation-of-care decisions such as life-sustaining treatment (LST) discussions

The predefined threshold of DeepCARS ≥ 91 has been shown in prior work to balance sensitivity and specificity for predicting in-hospital cardiac arrest. In our institution, this cutoff is also embedded in clinical workflow as an actionable alert [9,10,11], providing a pragmatic criterion for identifying patients at heightened risk. However, how clinicians interpret and respond to such alerts in real-world settings—beyond mortality prediction—remains unclear. However, high predictive accuracy in a model does not automatically translate to improved clinical outcomes. The real-world utility of such a system critically depends on how front-line clinicians interpret and act upon its alerts within their complex daily workflow. This raises crucial questions: Do physicians trust the AI-generated recommendations? How are these alerts prioritized amidst numerous other clinical tasks? Answering these questions is essential for understanding the true clinical impact of AI tools beyond their algorithmic performance. This study was conducted during the nationwide healthcare workforce crisis, which provided a unique context to evaluate real-world physician responses under resource constraints, although the development of DeepCARS itself preceded this crisis.

The present study aimed to evaluate physicians’ responses to DeepCARS alerts (≥91) in general ward patients at a tertiary hospital in Korea, during a nationwide healthcare workforce crisis when medical staffing was markedly reduced. Specifically, we examined factors associated with distinct clinical responses, including consultation, ICU transfer, or LST decision-making. By doing so, this study provides insight into the broader clinical utility of AI-based EWS and highlights how patient characteristics, comorbidities, and physiological derangements interact with AI alerts to guide individualized care decisions in practice.

This study evaluated patients admitted to a tertiary university hospital during a nationwide healthcare workforce crisis, during which medical staffing was significantly reduced. Under these constrained conditions, the real-world clinical application of the DeepCARS system, an AI-based early warning tool originally developed to predict in-hospital cardiac arrest, was evaluated. The findings of this study suggest that the DeepCARS system may not only predict cardiac events but also may provide a practical threshold for initiating physician-driven clinical interventions in high-risk ward patients. To our knowledge, this is the first study in Korea to examine the real-world implementation of DeepCARS in general wards during a period of medical system disruption.

Perhaps the most striking finding of this study was that patients selected for active intervention by clinicians had a dramatically higher mortality rate than those managed with observation alone, despite all patients having a high-risk DeepCARS score. This highlights the profound predictive power of comprehensive clinical judgment, which appears to surpass that of the AI alert itself. As the DeepCARS algorithm is based on quantitative data like vital signs, it has an inherent limitation in that it does not account for crucial prognostic factors such as a patient’s overall functional status, ability to walk, frailty, sarcopenia, or dementia. These non-quantitative factors are often the first things clinicians assess. A high DeepCARS score in a patient with good functional status may have been interpreted by clinicians as a reversible issue, warranting only observation. Conversely, a similar score in a frail, sarcopenic, or cognitively impaired patient may have prompted a decision for intervention, often a limitation of care (LST decision), because the clinician judged the patient’s potential for meaningful recovery to be low. Therefore, our findings affirm that AI-based early warning systems are best viewed as valuable support tools that alert clinicians, rather than as replacements for clinical assessment. The ultimate, nuanced decision-making—which weighs the patient’s complete clinical picture—remains firmly in the domain of human expertise. Future iterations of AI warning systems should aim to incorporate these critical functional and qualitative variables to improve their clinical utility.

Although all patients met the predefined high-risk threshold of DeepCARS ≥ 91, the mean score was significantly higher among those who received active intervention compared with those managed with observation alone. This indicates that clinicians should consider the absolute value of the score, and not just whether the alert is triggered, when determining the need for intervention. Notably, although our institution pragmatically used a cutoff of ≥91 as the actionable threshold, ROC analysis in this study suggested that ≥94 may improve discrimination for predicting physician intervention. This discrepancy underscores the need to continuously refine AI-based thresholds in real-world settings and highlights the critical balance between sensitivity and specificity when integrating AI tools into clinical workflows.

Beyond mortality prediction, our analysis indicated that DeepCARS scores ≥ 94 were associated with a higher likelihood of receiving active intervention, suggesting that this threshold may serve as a useful marker to guide clinical decisions. Comorbidities also appeared to influence physicians’ responses. Chronic liver disease showed an independent association with intervention, while hemato-oncologic comorbidities, although not statistically significant in the multivariate model, were more frequently observed among patients receiving active care. Taken together, these findings suggest that both disease-specific vulnerabilities and physiologic derangements reflected in the DeepCARS score could be related to physicians’ clinical responses in real-world practice.

Interestingly, the patterns of physician decision-making observed in this study were consistent with both baseline characteristics and regression analyses. Patients admitted to medical departments were more often directed toward LST decisions rather than ICU transfer, reflecting the clinical context in which these patients were typically managed. Beyond admission type, specific comorbidities also appeared to shape the direction of response: those with cardiovascular diseases were more frequently considered for ICU transfer compared with consultation alone, while patients with hemato-oncologic conditions were more likely to proceed either to ICU transfer or to LST decisions. Physiological derangements further contributed to these choices, with elevated respiratory rate emerging as a common signal prompting both ICU transfer and LST discussions. Together, these findings suggest that physician responses following high-risk alerts are not determined by a single factor but reflect an interplay between underlying vulnerabilities, the admitting department, and acute physiological changes.

About 40% of patients in this study with high DeepCARS scores were managed conservatively without active intervention. Notably, these patients showed relatively low mortality, indicating that in some cases alerts did not necessitate escalation. This suggests that certain alerts may reflect transient or reversible abnormalities, and that clinicians appropriately exercised discretion in filtering meaningful signals from noise. These observations emphasize that while AI-based systems can provide valuable guidance, their optimal role is to support, rather than replace, individualized clinical judgment.

Importantly, this study was conducted during a nationwide healthcare workforce crisis, which provided a unique stress test for the DeepCARS system. Physician decision-making under severe resource constraints may differ from usual practice, yet the emergence of consistent and interpretable patterns of response in this setting suggests the robustness of AI-based alerts even under strained healthcare environments. This context strengthens the generalizability of our findings, as it illustrates how AI tools may function as decision-support mechanisms during both routine practice and periods of systemic disruption.

These findings underscore the importance of viewing AI-generated alerts as part of a broader decision-making process, rather than as stand-alone triggers. Building on this perspective, further research is needed to define optimal thresholds for AI-triggered responses, clarify the role of clinical judgment in these responses, and evaluate the applicability of early warning systems such as DeepCARS in different hospital environments and patient populations. Prospective multicenter studies that incorporate response time, patient outcomes, and clinician perspectives will be essential to establish standardized pathways for effectively integrating AI alerts into routine clinical workflows [12].

This study has several limitations. First, it was conducted at a single tertiary-care hospital during a nationwide healthcare workforce crisis, which may limit the generalizability of the findings to other settings or to periods of normal staffing. The reduced availability of physicians during this period may have influenced both the timeliness and the nature of responses to DeepCARS alerts. Second, the retrospective design introduces inherent limitations, including the potential for selection bias and unmeasured confounding. Third, clinical responses following DeepCARS alerts—such as intervention, ICU transfer, or LST decisions—were not standardized and may have varied depending on physician judgment and clinical circumstances. Finally, due to the retrospective design of this study, we were unable to collect precise data on the time interval between the DeepCARS alert and subsequent clinical outcomes, such as ICU transfer or in-hospital death. This is a notable limitation, as this temporal data could have provided further insight into whether the timing of a patient’s deterioration (e.g., early vs. late in the hospital stay) influenced the type of clinical response, such as the decision between ICU transfer and an LST decision.

In conclusion, this study demonstrated the effectiveness of the DeepCARS system as a clinical decision-support tool during a healthcare crisis. Its primary value was serving as a reliable trigger that prompted physicians to make critical and timely clinical decisions—ranging from ICU transfer to end-of-life care planning—for appropriately identified high-risk patients.

The following abbreviations are used in this manuscript:

Conceptualization, H.J. and K.L.; methodology, H.J., W.Y. and S.H.; software, S.H.; validation, K.L.; formal analysis, H.J. and W.Y.; investigation, H.J.; resources, S.H.; data curation, H.J. and S.H.; writing—original draft preparation, H.J. and K.L.; writing—review and editing, H.J. and K.L.; visualization, H.J., W.Y., S.H. and K.L.; supervision, K.L.; project administration, H.J., W.Y. and K.L.; funding acquisition, K.L. All authors have read and agreed to the published version of the manuscript.

The study protocol was approved by the Institutional Review Board (IRB) of our institute, which waived the requirement for patient informed consent because of the retrospective and observational nature of the study (IRB No. 2504-011-150). All study procedures conformed to the Declaration of Helsinki 2024 and the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.

Patient consent was waived due to the retrospective and observational nature of the study.

The data presented in this study are available upon request from the corresponding author. The data are not publicly available due to privacy regulations.

The authors declare no conflicts of interest.

This work was supported by clinical research grant from Pusan National University Hospital in 2025.