Aplicação de algoritmos de machine learning para análise e predição de desfechos clínicos em pacientes com covid-19

Abstract

The COVID-19 pandemic imposed substantial challenges on healthcare systems worldwide, particularly due to the rapid clinical deterioration observed in many patients and the urgent need to allocate ventilatory support and intensive care resources efficiently, highlighting a gap in objective tools capable of assisting early risk stratification and outcome prediction. In this context, Machine Learning (ML) models, such as decision trees, offer the potential to support clinical decision-making by analyzing multiple variables simultaneously and generating easily interpretable predictive structures. This study aimed to apply ML algorithms to clinical data from hospitalized patients with COVID-19 to predict relevant clinical outcomes and identify factors associated with disease severity. This retrospective, quantitative study used secondary data from the public Severe Acute Respiratory Infection (SARI) database of Londrina, Paraná, Brazil, which initially contained 15,655 records of hospitalized patients between January 2021 and February 2022. After data cleaning and preprocessing, 5,704 records were used for the outcome prediction model and 7,182 for complementary analyses. Data were analyzed using the WEKA software (v.3.8.6) with the J48 decision tree algorithm and cross-validation, following attribute selection and removal of redundant or incomplete variables. The generated models achieved approximately 80% accuracy. Type of ventilatory support emerged as the most relevant predictor across analyses, followed by ICU admission, preexisting heart disease, age, and hospital type. Results showed that invasive mechanical ventilation and ICU admission were strongly associated with mortality, whereas non-invasive ventilation or absence of ventilatory support were associated with recovery. Advanced age substantially increased mortality risk among patients receiving invasive ventilation, and hospital type influenced outcomes among those not admitted to the ICU, with higher mortality in public hospitals. The study concludes that decision tree models are effective for identifying predictors of severity in hospitalized COVID-19 patients, providing clinically interpretable structures with potential application for risk stratification, resource allocation, and management of critical care pathways, especially in contexts of healthcare overload such as that experienced during the pandemic.