Predicting public Intensive care unit mortality and hospitalization using Data: An evaluation of Brazil's Largest...

To assess the SRAG dataset's potential for modeling COVID-19 mortality and LOS-ICU, identify key data gaps, and support the development of predictive tools for ICU planning in future outbreaks.

The SRAG dataset was split into training and test sets, followed by a hybrid feature variable selection strategy, which applied the XGBoost classifier and subsequent inclusion of comorbidities. Mortality prediction employed six supervised learning algorithms, while LOS-ICU was modeled using five regression techniques. Evaluation metrics included Receiver Operating Characteristic (ROC) curves, F1-score, sensitivity, precision, and specificity for mortality classification; and Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R² for LOS-ICU. Models were tuned to address data imbalance and improve generalizability across the dataset.

A total of 365,572 ICU inpatients were analyzed. Predictive models estimated mortality at around 51%, compared to an actual observed rate of ∼ 54%, with good discriminatory ability (AUC between 0.73 and 0.85). The average ICU stay was 11.6 days. Boosting models achieved the best performance for mortality prediction, with XGBoost reaching AUC = 0.85 and F1-score = 0.80, supporting early identification of high-risk patients. LOS-ICU regressions yielded MAE = 4-6 days and R² = 0.35-0.37, indicating a predicted ICU stay of 11.6 ± 6 days and reflecting limited explanatory power due to missing clinical biomarkers. Mortality predictions were robust for triage support, while ICU stay estimates, though less precise, remain useful for operational planning.

This research pioneered the use of a large-scale, highly granular public dataset and leveraged it for mortality prediction in COVID-19 patients in public ICUs. Given the unprecedented scale, academic and managerial improvements were brought to the field, offering generalizable solutions to support ICU triage and early warnings during outbreaks. Addressing the gaps of SRAG could enhance its utility for predictive modeling, especially in LOS-ICU modeling, enabling more accurate results.

Predictive modeling based on real-world ICU data can help healthcare decision-makers anticipate ICU demand, optimize triage, and allocate critical care resources with greater equity and timeliness during future epidemiological crises, ultimately improving patient outcomes.