Prediction of in-hospital death among patients admitted to a tertiary care hospital over the first ten years: A machine learning approach

Edel Rafael Rodea-Montero^1,2Brenda Jesús Rodríguez-Alcántar²Dagoberto Armenta-Medina^1*

• ¹Centro de Investigación e Innovación en Tecnologías de la Información y Comunicación, Aguascalientes, Mexico
• ²Hospital Regional de Alta Especialidad del Bajío, Guanajuato, Mexico

Purpose: To describe the pre-and postadmission characteristics of hospitalized patients in a tertiary care hospital and to adjust machine learning models capable of predicting and identifying the factors that are associated with and have a greater prognostic value for in-hospital death. Materials and methods: This was a retrospective study based on data from patients who were discharged from a Mexican tertiary care hospital during its first ten years of operation (2007-2016). Preadmission characteristics were analyzed using descriptive statistics. Comparison tests (Mann‒ Whitney U) and association tests (chi-square) were applied according to the absence or presence of in-hospital death. Multivariate models (logistic regression, random forest and XGBoost) were fitted. Their ROC curves were compared using the DeLong test, and performance metrics were evaluated. Results: In total, 55,253 hospital discharges were considered, only 45,011 (0-101 years) had complete data, and the rate of in-hospital death was 4.17%. In total, 70% of the data were used for training and 30% for testing. Two-to-two comparisons between areas under the curve (AUCs) revealed that XGBoost (AUC = 0.9162) outperformed logistic regression (AUC = 0.9036) and random forest (AUC = 0.8978) (p value < 0.0001 in both cases). XGBoost had a sensitivity of 87%, specificity of 81.3% and balanced efficiency of 84.2%. The most relevant predictive factors were medical service that performed the admission, number of conditions, origin of the outpatient consultation of the hospital, and the main condition diagnosed at admission according to the ICD-10, age, month of admission, and day of the week of admission. Conclusions: Owing to its ability to capture complex patterns, the XGBoost model makes it possible to identify patients with a relatively high risk of in-hospital death using the data available at hospital admission. This constitutes a support tool for decision-making, helping to determine which patients require closer monitoring and follow-up during their hospital stay to improve the quality of medical care.