Background/Objectives: Data availability can affect the performance of AI-based early warning scores (EWSs). This study evaluated how the extent of missing data and imputation strategies influence the predictive performance of the VitalCare–Major Adverse Event Score (VC-MAES), an AI-based EWS that uses last observation carried forward and normal-value imputation for missing values, to forecast clinical deterioration events, including unplanned ICU transfers, cardiac arrests, or death, up to 6 h in advance. Methods: We analyzed real-world data from 6039 patient encounters at Keimyung University Dongsan Hospital, Republic of Korea. Performance was evaluated under three scenarios: (1) using only vital signs and age, treating all other variables as missing; (2) reintroducing a full set of real-world clinical variables; and (3) imputing missing values drawn from a distribution within one standard deviation of the observed mean or using Multiple Imputation by Chained Equations (MICE). Results: VC-MAES achieved the area under the receiver operating characteristic curve (AUROC) of 0.896 using only vital signs and age, outperforming traditional EWSs, including the National Early Warning Score (0.797) and the Modified Early Warning Score (0.722). Reintroducing full clinical variables improved the AUROC to 0.918, whereas mean-based imputation or MICE decreased the performance to 0.885 and 0.827, respectively. Conclusions: VC-MAES demonstrates robust predictive performance with limited inputs, outperforming traditional EWSs. Incorporating actual clinical data significantly improved accuracy. In contrast, mean-based or MICE imputation yielded poorer results than the default normal-value imputation, potentially due to disregarding the “informative presence” embedded in missing data patterns. These findings underscore the importance of understanding missingness patterns and employing imputation strategies that consider the decision-making context behind data availability to enhance model reliability.

Objective: The heterogeneous nature of sepsis renders determining its underlying causes difficult, which may delay diagnosis and intervention. VitalCare-SEPsis Score (VC-SEPS) is a deep learning-based algorithm that predicts sepsis and monitors patient conditions based on electronic medical record data. However, few studies have prospectively compared medical artificial intelligence software algorithms and traditional scoring systems to predict sepsis. This prospective observational study attempted to validate the predictive performance and risk stratification of VC-SEPS for early prediction of sepsis. Methods: In this prospective observational study, we collected electronic medical record data from 6,797 patients hospitalised at Keimyung University Dongsan Hospital, Daegu, South Korea. The final version of the analysed set included 6,455 patients, 325 of whom were diagnosed with sepsis. Results: The area under the receiver operating characteristic curve of VC-SEPS was 0.880, indicating its superiority over traditional scoring systems. The algorithm performance showed a consistent trend within 24 hours. On patients’ initial admission, the VC-SEPS was associated with the risk of developing sepsis, and the score accurately predicted sepsis by an average of 68.05 min compared with diagnosis time by an operational definition of sepsis. Discussion: VC-SEPS could assist medical staff with early diagnosis and intervention in clinical practice by providing a sepsis risk score. Prompt recognition assisting recognition can significantly help shorten the time between recognition and intervention in clinical decision-making processes. Conclusion: This study suggests that using a clinical decision support system can help improve hospital workflows as well as the quality of medical care.

Background: Acute deterioration of patients in general wards often leads to major adverse events (MAEs), including unplanned intensive care unit transfers, cardiac arrest, or death. Traditional early warning scores (EWSs) have shown limited predictive accuracy, with frequent false positives. We conducted a prospective observational external validation study of an artificial intelligence (AI)-based EWS, the VitalCare - Major Adverse Event Score (VC-MAES), at a tertiary medical center in the Republic of Korea. Methods: Adult patients from general wards, including internal medicine (IM) and obstetrics and gynecology (OBGYN)—the latter were rarely investigated in prior AI-based EWS studies—were included. The VC-MAES predictions were compared with National Early Warning Score (NEWS) and Modified Early Warning Score (MEWS) predictions using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and logistic regression for baseline EWS values. False-positives per true positive (FPpTP) were assessed based on the power threshold. Results: Of 6,039 encounters, 217 (3.6%) had MAEs (IM: 9.5%, OBGYN: 0.26%). Six hours prior to MAEs, the VC-MAES achieved an AUROC of 0.918 and an AUPRC of 0.352, including the OBGYN subgroup (AUROC, 0.964; AUPRC, 0.388), outperforming the NEWS (0.797 and 0.124) and MEWS (0.722 and 0.079). The FPpTP was reduced by up to 71%. Baseline VC-MAES was strongly associated with MAEs (P<0.001). Conclusions: The VC-MAES significantly outperformed traditional EWSs in predicting adverse events in general ward patients. The robust performance and lower FPpTP suggest that broader adoption of the VC-MAES may improve clinical efficiency and resource allocation in general wards.

Background: In hospitals, Code Blue is an emergency that refers to a patient requiring immediate resuscitation. Over 85% of patients with cardiopulmonary arrest exhibit abnormal vital sign trends prior to the event. Continuous monitoring and accurate interpretation of clinical data through artificial intelligence (AI) models can contribute to preventing critical events. Objective: This study aims to evaluate changes in clinical outcomes following the use of VitalCare (Major Adverse Event Score and Mortality Score), which is an AI-based early warning system, and to validate the performance of the algorithm. Methods: A retrospective analysis was conducted by extracting electronic health record data, using a total of 30,785 inpatient cases from general wards and intensive care units. A comparative analysis was performed by setting a 3-month period before and after the system implementation. For clinical evaluation, we measured the incidence rates of Code Blue and adverse events, the proportion of prolonged hospitalization, and the frequency of early interventions. The area under the receiver operating characteristic curve (AUROC) was calculated to assess the performance of the algorithm. Results: This study demonstrated that, following the implementation of VitalCare, there was a 24.97% reduction in major events such as Code Blue (P=.004) and the proportion of prolonged hospitalization in general wards (P<.05), along with a significant increase in the rate of early interventions. The model performance exhibited superior outcomes compared with traditional scoring systems, with a Major Adverse Event Score AUROC of 0.865 (95% CI 0.857‐0.873) and Mortality Score AUROC of 0.937 (95% CI 0.931‐0.944). Conclusions: A well-developed AI-based model that provides high predictive power can contribute to the prevention of major in-hospital events by providing early predictive information to clinicians. Additionally, it plays a crucial role in effectively addressing unmet needs and challenges in terms of human resources and practical procedures.

The timely detection of ward deterioration—including unplanned intensive care unit (ICU) transfer, cardiac arrest, death, and sepsis—remains an unmet need. Although rule-based early warning scores and newer machine-learning models have been introduced, their clinical adoption is limited owing to challenges such as low predictive performance, excessive false alarms, and lack of generalizability. This study aimed to develop two deep-learning models for patients in general wards using a bidirectional long short-term memory neural network architecture: (i) the VitalCare-Major Adverse Event Score (VC-MAES), which predicts clinical deterioration events (CDEs)— unplanned ICU transfer, cardiac arrest, or in-hospital death—within 6 h and (ii) the VitalCare-SEPsis Score (VC-SEPS), which predicts sepsis onset within 4 h. Additionally, we sought to externally validate the performance of the models in an independent cohort. This study was conducted in two sequential phases. First, the VC-MAES and VC-SEPS models were developed using a large retrospective cohort from Yonsei Severance Hospital, Seoul, Republic of Korea. Second, external validation was performed in a single-center cohort at National Health Insurance Service Ilsan Hospital (NHIS Ilsan Hospital), Ilsan, Republic of Korea. Both algorithms incorporated patient age, vital signs, laboratory results, and the Glasgow Coma Scale scores. VC-MAES performance was compared with the Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS), whereas VC-SEPS performance was compared with the Sequential Organ Failure Assessment (SOFA), quick SOFA (qSOFA), and NEWS, using the area under the receiver operating characteristic curve (AUROC) as the primary performance metric. The derivation cohort comprised 357,009 adult general-ward admissions at Yonsei Severance Hospital (2013–2017), and the external validation cohort included 22,073 admissions at NHIS Ilsan Hospital (2017). In the external validation cohort, the VC-MAES predicted CDEs within 6 h with an AUROC of 0.918 (95% confidence interval [CI], 0.909–0.927), outperforming the MEWS (0.834; 95% CI, 0.820–0.849) and NEWS (0.883; 95% CI, 0.869– 0.896). VC-SEPS predicted sepsis onset within 4 h, with an AUROC of 0.941 (95% CI, 0.934–0.947), surpassing the SOFA (0.559; 95% CI, 0.546–0.571), qSOFA (0.687; 95% CI, 0.671–0.704), and NEWS (0.767; 95% CI, 0.748–0.785). Both models maintained AUROC values above 0.86 across all age and sex categories. The VC-MAES and VC-SEPS outperformed conventional early warning scores in predicting clinical deterioration events and sepsis. These models can enable earlier and more precise interventions, enhancing patient care and optimizing resource use, ultimately leading to better patient outcomes without overwhelming healthcare providers.

Background: This retrospective study introduces an AI-driven VitalCare-Major Adverse Event Score (VC-MAES) developed to predict major in-hospital adverse events and determine optimal cutoff thresholds. VC-MAES was originally developed to predict a composite outcome including unplanned intensive care unit (ICU) transfer, in-hospital cardiac arrest, and mortality; the present study evaluates its performance for the composite of unplanned ICU transfer and in-hospital cardiac arrest. Methods: Patients aged ≥19 years in Yongin Severance Hospital between 1 March 2020, and 31 December 2022, were included. The primary outcome was clinica l deterioration, defined as unplanned ICU transfer or in-hospital cardiac arrest. Secondary outcomes included model performance metrics (area under the receiver operating characteristic (ROC) curve, sensitivity, specificity, F1 score) and optimal alarm frequency (number of alarms per 100 patient-days) across cutoff determination methods (Youden’s index, F1 score, Euclidean distance, and alarm-based approach); they were used to determine optimal cutoff values. 20 July 2026. Results: VC-MAES achieved an area under the ROC curve of 0.895(full evaluable sample) at 6-h intervals, outperforming traditional early warning systems. It exhibited improved sensitivity (0.5853) compared to NEWS (0.3775) and MEWS (0.3049) while maintaining a specificity exceeding 0.95, balancing alarm frequency and predictive accuracy (all figures based on each model’s own full evaluable sample: VC-MAES, n = 7425 events; NEWS, n = 3494; MEWS, n = 3198). On the common prediction time-point sample (n = 1678 events), the alarm-based cutoff yielded a sensitivity of 0.5244 for VC-MAES versus 0.2271 for NEWS and 0.1740 for MEWS, with alarm frequencies of 9.67, 7.07, and 5.58 per 100 patient-days, respectively. When compared on a common set of prediction time-points, VC-MAES (AUROC 0.832 on the common sample, 95% CI: 0.822–0.842) significantly outperformed NEWS (AUROC 0.743, 95% CI: 0.732–0.756; DeLong Z = −14.19, p < 0.0001) and MEWS (AUROC 0.706, 95% CI: 0.693–0.720; DeLong Z = −18.22, p < 0.0001). ROC-based methods, such as Youden’s index and Euclidean distance, were deemed unsuitable for practical clinical use. Conclusions: Conventional ROC-based cutoff selection methods may not be suitable for real-world clinical implementation due to excessive alarm burden. An alarm-based cutoff approach provides a more clinically applicable strategy by balancing predictive performance with manageable alarm frequency. These findings highlight the importance of incorporating clinical usability into cutoff determination for AI-based early warning systems.