Your browser doesn't support javascript.
Clinical profile and outcomes of patients with COVID-19 respiratory failure requiring renal replacement therapy: a machine learning analysis of 13,576 patients
Critical Care Conference: 42nd International Symposium on Intensive Care and Emergency Medicine Brussels Belgium ; 27(Supplement 1), 2023.
Article in English | EMBASE | ID: covidwho-2314604
ABSTRACT

Introduction:

Acute kidney injury (AKI) is a frequent and severe complication of COVID-19 infection in ICU patients. We propose a structured data-driven methodology and develop a model to predict the use of renal replacement therapy for patients on respiratory support with Covid-19 in 126 ICUs from 42 Brazilian hospitals. Method(s) Adult ICU patients (March 2020-December 2021) with confirmed SARS-CoV-2 infection and need of ventilatory support at D1 admission in the ICU. Main outcome was the need of RRT. We estimated 3 prediction models Logistic Regression (LR), Random Forest (RF) and XGB Boosting. Models were derived in the training set and evaluated in the test set following an 80/20 split ratio, and models' parameters were selected using fivefold cross-validation. We evaluated and selected the best model in terms of discrimination (AUC) and calibration (Brier's Score). Variable importance was estimated for each predictor variable. Result(s) 13,575 ICU patients with need of respiratory support, of which 1828 (14%) needed RRT. ICU and hospital mortality were respectively 15.7%, 20.3% (non-RRT) and 54.3%, 69% (RRT). Mean age was 63.9 and 55.3 years (RRT vs non-RRT). Mean ICU LOS was 27.8 vs. 12 days, in RRT vs non-RRT. RF and XGB models both showed higher discrimination performance compared to LR (95% confidence interval [95% CI] 0.84 [0.81-0.85] and 0.83 [0.80-0.85] vs 0.78 [0.75-0.80]). RF and XGB models presented similar calibration (Brier's Score ([95% CI] 0.09 [0.09- 0.10] and 0.09 [0.09-0.10]), also better than in LR (0.11 [0.10-0.12]). The final model (RF) showed no sign of under or overestimation of predicted probabilities in calibration plots. Conclusion(s) The need of RRT among patients on respiratory support diagnosed with Covid-19 was accurately predicted through machine learning methods. RF and XGB based models using data from general intensive care databases provides an accurate and practical approach for the early prediction of use of RRT in COVID-19 patients.
Keywords

Full text: Available Collection: Databases of international organizations Database: EMBASE Type of study: Prognostic study Language: English Journal: Critical Care Conference: 42nd International Symposium on Intensive Care and Emergency Medicine Brussels Belgium Year: 2023 Document Type: Article

Similar

MEDLINE

...
LILACS

LIS


Full text: Available Collection: Databases of international organizations Database: EMBASE Type of study: Prognostic study Language: English Journal: Critical Care Conference: 42nd International Symposium on Intensive Care and Emergency Medicine Brussels Belgium Year: 2023 Document Type: Article