Aplicação técnica de aprendizado de máquina na predição de mortalidade hospitalar em pacientes acometidos por AVC hemorrágico - um estudo de caso MIMIC-IV

Abstract

This study aims to develop machine learning-based predictive models for predicting deaths in hospitalized patients suffering from hemorrhagic stroke, given the high mortality rate associated with the disease, thus enabling treat- ment to prevent fatalities. To this end, results from tests collected within the first 24 hours after admission to the healthcare unit were used. The MIMIC-IV database was used for this work. In this database, 2,976 patients over 18 years of age who had suffered hemorrhagic stroke were selected. Of this total, 1,367 died, with 513 of them dying in the hospital setting. Empirical tests were performed with supervised machine learning algorithms, in which XGBoost and LightGBM were selected for further investigation due to their superior performance based on the area under the curve (AUC). Similarly, the calibration curves of the resulting models were analyzed, in addition to the importance of each attribute in the models’ decisions, using the SHAP technique, which demonstrated that the final indicator of the Glasgow Coma Scale is the most impactful in predicting hospital mortality. Both XGBoost and LightGBM showed a high capacity for class discrimination, AUC = 0.915 and AUC = 0.911, respectively. The analysis of the calibration curves showed that the models do not tend to overestimate or underestimate risks significantly, which consequently guarantees their usefulness for decision-making in the hospital environment. Likewise, the clinical decision curves showed that both models present an associated cli- nical benefit, mainly in the low and medium probabilistic thresholds. In turn, the Precision-Recall curves demonstrated the relationship between precision and recall for the unbalanced set of tests. Both XGBoost and LightGBM have shown significant results for the early identification of deaths in patients with hemorrhagic stroke. This contributes to clinical decision-making by using laboratory data collected from patients, ensuring risk stratification and the potential to reduce future complications.