RESUMO
OBJECTIVE: Severity of illness scores used in critical care for benchmarking, quality assurance and risk stratification have been mainly created in high-income countries. In low and middle-income countries (LMICs), they cannot be widely utilized due to the demand for large amounts of data that may not be available (e.g. laboratory results). We attempt to create a new severity prognostication model using fewer variables that are easier to collect in an LMIC. SETTING: Two intensive care units, one private and one public, from São Paulo, Brazil PATIENTS: An ICU for the first time. INTERVENTIONS: None. MEASUREMENTS AND MAINS RESULTS: The dataset from the private ICU was used as a training set for model development to predict in-hospital mortality. Three different machine learning models were applied to five different blocks of candidate variables. The resulting 15 models were then validated on a separate dataset from the public ICU, and discrimination and calibration compared to identify the best model. The best performing model used logistic regression on a small set of 10 variables: highest respiratory rate, lowest systolic blood pressure, highest body temperature and Glasgow Coma Scale during the first hour of ICU admission; age; prior functional capacity; type of ICU admission; source of ICU admission; and length of hospital stay prior to ICU admission. On the validation dataset, our new score, named SEVERITAS, had an area under the receiver operating curve of 0.84 (0.82 - 0.86) and standardized mortality ratio of 1.00 (0.91-1.08). Moreover, SEVERITAS had similar discrimination compared to SAPS-3 and better discrimination than the simplified TropICS and R-MPM. CONCLUSIONS: Our study proposes a new ICU mortality prediction model using simple logistic regression on a small set of easily collected variables may be better suited than currently available models for use in low and middle-income countries.