Abstract
Background and Objectives: Timely recognition of deteriorating ward patients is critical to prevent adverse outcomes. The Deep learning-based Cardiac Arrest Risk Score (DeepCARS), an AI-based early warning system developed in Korea, has demonstrated high sensitivity and specificity, but its impact on real-world physician decision-making remains unclear, especially under healthcare resource constraints. Materials and Methods: We retrospectively analyzed 830 adult ward patients (March 2024-February 2025) who triggered DeepCARS alerts (score ≥ 91) at a tertiary hospital during a nationwide workforce shortage. Physician responses were classified as active intervention (ICU transfer, life-sustaining treatment [LST] decision, or specialty consultation) versus observation. Results: Among patients with DeepCARS ≥ 91, 58.9% received active intervention, with higher in-hospital mortality compared with those observed only (34.8% vs. 9.7%). ROC analysis suggested a cutoff of ≥94 for better intervention discrimination (AUC = 0.708). In multivariable analysis, DeepCARS ≥ 94 (OR 3.52) and chronic liver disease (OR 1.78) independently predicted active intervention. Multinomial analysis showed that patients admitted to medical departments were more often directed toward LST decisions rather than ICU transfer. Hemato-oncologic comorbidities were associated with both ICU transfer and LST decisions, while elevated respiratory rate consistently predicted either ICU transfer or LST discussions. Conclusions: DeepCARS alerts effectively triggered physician-driven decisions regarding ICU transfer and end-of-life care during a healthcare crisis. However, the ultimate clinical responses were shaped by comprehensive clinical judgment that integrated AI-generated risks with patient-specific factors, such as functional status and frailty, not captured by the algorithm. This underscores the indispensable role of individualized clinical assessment in interpreting and acting upon AI-based alerts in high-risk ward patients.