Algoritmo de predicción de sepsis basado en machine learning.
Loading...
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
En este proyecto de memoria se abordó la grave problemática de la sepsis, una de las condiciones que posee una tasa de mortalidad del 19,8 % a nivel mundial, y especialmente delicada en pacientes internados en unidades críticas. Ante la elevada tasa de letalidad y lo rápido que progresa esta condición, su detección temprana y atención inmediata son fundamentales para asegurar la mejora del paciente, por lo cual el objetivo del proyecto es desarrollar un algoritmo predictivo de sepsis basado en Machine Learning. Para lograrlo, se realizó una revisión bibliográfica para identificar los principales factores de riesgo, síntomas de la condición y modelos de aprendizaje automático ya utilizados, junto con las variables que poseen mayor incidencia en el desempeño de estos modelos. Utilizando la información recolectada, se procedió a avanzar en el desarrollo del modelo trabajando con los registros de la base de datos eICU CRD. Se preparó la base de datos seleccionando una cohorte y las variables más significativas según la bibliografía para, posteriormente, aplicar el preprocesamiento, que abarcó el análisis exploratorio de los datos, el manejo de datos faltantes, la eliminación de variables con alta correlación, la codificación de variables categóricas y el escalado de variables numéricas, el manejo de outliers, la gestión del desbalanceo de la clase objetivo y la selección de características según su nivel de importancia. Finalmente, con la base de datos preprocesada se aplicaron y evaluaron distintos modelos, tales como Regresión Logística, Support Vector Machine, Random Forest, XGBoost, LightGBM, CatBoost y una Red Neuronal de tipo Perceptrón Multicapa (MLP), a los cuales se le aplicó la optimización de sus hiperparámetros mediante Optuna. Al evaluar su desempeño se identificó que los modelos de potenciadores de gradiente poseen un desempeño superior, destacado el modelo de LightGBM, con un desempeño en AUC de 0,88. Además, se validó la interpretabilidad del modelo mediante el análisis SHAP de características, evidenciando como las variables más influyentes en la predicción concuerdan con síntomas o marcadores característicos de la condición de sepsis.
In this thesis project, we addressed the serious issue of sepsis, one of the conditions with a global mortality rate of 19.8%, and which is particularly critical among patients admitted to intensive care units. Given the high lethality rate and the speed with which this condition progresses, early detection and immediate intervention are essential to improving patient outcomes. For this reason, the project's objective is to develop a Machine Learning based predictive algorithm for sepsis. To achieve this, a literature review was conducted to identify the main risk factors, characteristic symptoms, and previously used machine learning models, as well as the variables with the most significant impact on the performance of these models. Using the collected information, the model development proceeded using records from the eICU CRD database. The database was prepared by selecting an appropriate cohort and identifying the most relevant features based on the literature. The preprocessing stage included exploratory data analysis, handling of missing data, removal of highly correlated variables, encoding of categorical features, scaling of numerical variables, management of outliers, addressing target class imbalance, and feature selection based on importance. Finally, with the dataset fully preprocessed, various models were implemented and evaluated, including Logistic Regression, Support Vector Machine (SVM), Random Forest, XGBoost, LightGBM, CatBoost, and a Multi-layer Perceptron Network (MLP). The hyperparameters of these models were optimized using Optuna. The performance evaluation revealed that gradient boosting models exhibited superior results, with LightGBM standing out as the most effective, achieving an AUC of 0,88. In addition, the final model's interpretability was validated through a SHAP feature analysis, which showed that the most influential variables in the prediction aligned with characteristic symptoms or biomarkers of the sepsis condition.
In this thesis project, we addressed the serious issue of sepsis, one of the conditions with a global mortality rate of 19.8%, and which is particularly critical among patients admitted to intensive care units. Given the high lethality rate and the speed with which this condition progresses, early detection and immediate intervention are essential to improving patient outcomes. For this reason, the project's objective is to develop a Machine Learning based predictive algorithm for sepsis. To achieve this, a literature review was conducted to identify the main risk factors, characteristic symptoms, and previously used machine learning models, as well as the variables with the most significant impact on the performance of these models. Using the collected information, the model development proceeded using records from the eICU CRD database. The database was prepared by selecting an appropriate cohort and identifying the most relevant features based on the literature. The preprocessing stage included exploratory data analysis, handling of missing data, removal of highly correlated variables, encoding of categorical features, scaling of numerical variables, management of outliers, addressing target class imbalance, and feature selection based on importance. Finally, with the dataset fully preprocessed, various models were implemented and evaluated, including Logistic Regression, Support Vector Machine (SVM), Random Forest, XGBoost, LightGBM, CatBoost, and a Multi-layer Perceptron Network (MLP). The hyperparameters of these models were optimized using Optuna. The performance evaluation revealed that gradient boosting models exhibited superior results, with LightGBM standing out as the most effective, achieving an AUC of 0,88. In addition, the final model's interpretability was validated through a SHAP feature analysis, which showed that the most influential variables in the prediction aligned with characteristic symptoms or biomarkers of the sepsis condition.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Biomédico/a.
Keywords
Algoritmos, Predicciones, Sepsis, Machine learning