Clinical Validation of Vital-PICASO for Predicting Cardiac Arrest Within 24 Hours

Overview

The purpose of this retrospective study is to evaluate the clinical performance of Vital-PICASO, an artificial intelligence-based biological signal analysis software designed to predict the risk of in-hospital cardiac arrest within 24 hours using vital-sign data collected from adult general ward inpatients. Electronic medical record data from patients aged 19 years or older who were admitted to a general ward at Seoul National University Hospital will be retrospectively reviewed. Vital-sign variables include systolic blood pressure, diastolic blood pressure, heart rate, respiratory rate, body temperature, and, when available, oxygen saturation. Eligible data will be classified as cardiac arrest-positive or cardiac arrest-negative according to predefined reference-standard criteria. The blinded datasets will then be analyzed using Vital-PICASO, and the software-generated risk scores will be compared with the reference-standard classifications. Predictive performance will be evaluated separately using models that include oxygen saturation and models that do not include oxygen saturation.

This study is a retrospective, single-center, single-arm, superiority-confirmatory clinical performance study conducted at Seoul National University Hospital. Data Source and Study Population Electronic medical records will be reviewed to identify patients aged 19 years or older who were admitted to a general ward for at least 12 hours between January 1, 2023, and May 31, 2025. Data previously used in the development of the investigational software will be excluded. The clinical variables collected from eligible records include systolic blood pressure, diastolic blood pressure, heart rate, respiratory rate, body temperature, oxygen saturation when available, and the corresponding measurement times. Additional variables include the occurrence, timing, and location of in-hospital cardiac arrest, month and year of birth, sex, height and weight when available, do-not-resuscitate status, date of admission to the general ward, and clinical department. Data Selection Potentially eligible cardiac arrest-positive and cardiac arrest-negative records will be identified from the electronic medical record system. To reduce selection bias, candidate records will be numbered and randomly selected using statistical software. Cardiac arrest-positive data will include cases in which cardiac arrest occurred in a general ward and was documented in the electronic medical record, followed by resuscitation involving high-quality cardiopulmonary resuscitation, defibrillation, or advanced life support. For patients with multiple cardiac arrest events, only the first eligible event will be included. Vital-sign data from up to 72 hours before the cardiac arrest event will be collected. If the event occurred within 24 hours after admission to the general ward, all available vital-sign data before the event will be collected. Cardiac arrest-negative data will include patients who did not experience cardiac arrest or other predefined clinical deterioration during the general ward admission. Negative cases will be randomly selected from the same wards and from time periods comparable with those of the positive cases. Up to 72 hours of vital-sign data will be collected, with one eligible data segment selected per patient. The planned study dataset consists of 387 cases, including 65 cardiac arrest-positive cases and 322 cardiac arrest-negative cases. Two predefined analysis sets will be evaluated: 178 cases for the analysis including oxygen saturation and 209 cases for the analysis excluding oxygen saturation. Records may overlap between the two analysis sets. Reference Standard Establishment The initial positive or negative classification will be determined from the electronic medical record according to the predefined cardiac arrest criteria. To improve the reliability of the study dataset, a clinician with at least 1 year of clinical experience, blinded to the initial positive or negative classification, will independently review the screening number, reference time point, and relevant clinical information. The clinician's final classification will be used as the reference standard. Application of the Investigational Device The medical device operator will receive vital-sign datasets identified only by screening numbers and will be blinded to the reference-standard classification. Each dataset will be analyzed using Vital-PICASO to generate the predicted risk of cardiac arrest within 24 hours. The peak risk score, ranging from 0 to 100 percent, and the corresponding time will be recorded. Statistical Analysis The software-generated risk scores will be compared with the reference-standard classifications. The primary performance measures are the area under the receiver operating characteristic curve for: 1. the model including oxygen saturation; and 2. the model excluding oxygen saturation. The clinical performance of each model will be compared with the predefined reference performance value derived from conventional early warning scores. The performance criterion for each primary outcome will be met if the lower bound of the 95 percent confidence interval for the difference between the Vital-PICASO AUROC and the reference AUROC is greater than zero. The overall study will be considered successful if both primary outcomes meet their respective performance criteria. Secondary performance measures include the area under the precision-recall curve, precision, recall, F1 score, threshold-specific performance, performance according to prediction-time intervals, and cumulative-time performance. Because this study uses previously collected electronic medical record data, there will be no direct participant contact, additional examination, treatment, or change in clinical care.