Predicción de mortalidad en pacientes con ECV hospitalizados en UCI mediante machine learning y la base de datos MIMIC-III.
Loading...
Date
2026
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
En este proyecto se aborda la predicción de mortalidad en pacientes con enfermedades cardiovasculares (ECV) hospitalizados en UCI, utilizando la base de datos MIMIC-III v1.4 como fuente de información. Se definieron objetivos y alcances enfocados en una cohorte de pacientes UCI adultos con los principales diagnósticos ECV registrados en la base de datos (insuficiencia cardíaca, fibrilaciones auriculares, cardiopatías isquémicas, infartos agudos al miocardio, entre otras.).
La metodología de trabajo incluyó una revisión de literatura inicial, la exploración de MIMIC-III mediante consultas SQL y la construcción de cohortes bajo criterios de integridad (edad y tiempo mínimo en UCI). Posteriormente, se seleccionaron las variables clínicas más relevantes y se diseñó un pipeline en Python para el preprocesamiento, construcción del dataset y entrenamiento de los modelos de machine learning, donde luego, se utilizaron métricas estándar de desempeño para evaluar los modelos mediante un esquema que permitió analizar tanto su rendimiento predictivo como la estabilidad de los resultados frente a las particiones de los datos. Debido a que MIMIC-III es una base datos masiva, tanto los análisis exploratorios de la base de datos, como el desarrollo y evaluación de los modelos se realizaron utilizando un servidor local de la Universidad.
Finalmente, los modelos basados en gradient boosting (XGBoost, LightGBM y Catboost) obtuvieron el mejor rendimiento, alcanzando valores de AUC-ROC cercanos a 0,80 y superando en gran medida al modelo baseline Random Forest. El análisis de variables más influyentes en el rendimiento de los modelos mostró patrones coherentes con la fisiopatología cardiovascular, destacando parámetros bioquímicos, hematológicos y demográficos tales como los electrolitos principales, hemogramas o la edad. Esto refuerza el potencial del aprendizaje automático como herramienta de apoyo en la predicción de mortalidad en pacientes con ECV ingresados a UCI.
This project addresses the prediction of mortality in patients with cardiovascular diseases (CVD) admitted to the intensive care unit (ICU), using the MIMIC-III v1.4 database as the primary data source. The objectives and scope of the study were defined focusing on a cohort of adult ICU patients with the main cardiovascular diagnoses recorded in the database, including heart failure, atrial fibrillation, ischemic heart disease, acute myocardial infarction, among others. The methodology comprised an initial literature review, exploration of the MIMIC-III database through SQL queries, and cohort construction under integrity criteria such as age and minimum ICU length of stay. Subsequently, the most relevant clinical variables were selected, and a Python-based pipeline was designed for data preprocessing, dataset construction, and machine learning model training. Model performance was evaluated using standard evaluation metrics under a framework that enabled the assessment of both predictive performance and result stability across different data partitions. Given the large-scale nature of the MIMIC-III database, exploratory data analysis as well as model development and evaluation were conducted using a local server provided by the University. Finally, gradient boosting–based models (XGBoost, LightGBM, and CatBoost) achieved the best performance, reaching AUC-ROC values close to 0,80 and substantially outperforming the Random Forest baseline model. The analysis of the most influential variables revealed patterns consistent with cardiovascular pathophysiology, highlighting biochemical, hematological, and demographic parameters such as key electrolytes, blood count indicators, and age. These findings reinforce the potential of machine learning as a supportive tool for mortality prediction in ICU patients with cardiovascular diseases.
This project addresses the prediction of mortality in patients with cardiovascular diseases (CVD) admitted to the intensive care unit (ICU), using the MIMIC-III v1.4 database as the primary data source. The objectives and scope of the study were defined focusing on a cohort of adult ICU patients with the main cardiovascular diagnoses recorded in the database, including heart failure, atrial fibrillation, ischemic heart disease, acute myocardial infarction, among others. The methodology comprised an initial literature review, exploration of the MIMIC-III database through SQL queries, and cohort construction under integrity criteria such as age and minimum ICU length of stay. Subsequently, the most relevant clinical variables were selected, and a Python-based pipeline was designed for data preprocessing, dataset construction, and machine learning model training. Model performance was evaluated using standard evaluation metrics under a framework that enabled the assessment of both predictive performance and result stability across different data partitions. Given the large-scale nature of the MIMIC-III database, exploratory data analysis as well as model development and evaluation were conducted using a local server provided by the University. Finally, gradient boosting–based models (XGBoost, LightGBM, and CatBoost) achieved the best performance, reaching AUC-ROC values close to 0,80 and substantially outperforming the Random Forest baseline model. The analysis of the most influential variables revealed patterns consistent with cardiovascular pathophysiology, highlighting biochemical, hematological, and demographic parameters such as key electrolytes, blood count indicators, and age. These findings reinforce the potential of machine learning as a supportive tool for mortality prediction in ICU patients with cardiovascular diseases.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Biomédico/a.
Keywords
Enfermedades cardiovasculares, Mortalidad, Cuidados críticos