Predicción de IAAS en estancias UCI mediante base de datos MIMIC-III.
Loading...
Date
2026
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Las infecciones asociadas a la atención en salud (IAAS) constituyen un problema crítico de seguridad del paciente, especialmente en unidades de cuidados intensivos (UCI), donde la condición clínica grave, el uso de dispositivos invasivos y las estancias prolongadas incrementan la morbilidad, la mortalidad y los costos hospitalarios. A pesar de los programas de vigilancia, las estrategias tradicionales presentan limitaciones para la detección temprana de estos eventos en contextos de alta carga asistencial y grandes volúmenes de datos clínicos.
Este trabajo desarrolla y evalúa modelos de aprendizaje automático para la predicción temprana de IAAS en estancias UCI utilizando la base pública MIMIC-III v1.4, que contiene registros clínicos anonimizados de pacientes críticos del Beth Israel Deaconess Medical Center. Se seleccionó una cohorte de 28.763 estancias de pacientes adultos con permanencia en UCI ≥ 48 h, con una prevalencia de IAAS cercana al 3,3%. Las etiquetas de CLABSI, ITUAC y PVAP se derivaron de los criterios CDC/NHSN adaptados a MIMIC-III. Se construyó un conjunto de variables predictoras a partir de información demográfica, fisiológica, de laboratorio, microbiológica y de procedimientos, aplicando creación de tablas derivadas, limpieza de datos, tratamiento de rangos plausibles, preprocesamiento y manejo explícito del desbalance de clases.
Se implementaron modelos de referencia y clasificadores basados en Random Forest y XGBoost, combinados con ajuste de pesos de clase, submuestreo (NearMiss) y sobremuestreo (SMOTE). El desempeño se evaluó mediante validación cruzada estratificada y en un conjunto de prueba con prevalencia original, utilizando métricas adaptadas a eventos raros (AUPRC, F2, recall), además de AUROC. El mejor rendimiento se obtuvo con XGBoost baseline ponderado, alcanzando un AUPRC de 0,097 (frente a 0,033 de modelos triviales), un AUROC de 0,752, un recall de 0,593 y un F2 de 0,265, lo que evidencia una capacidad real, aunque moderada, para priorizar pacientes con mayor riesgo de IAAS en un escenario de baja prevalencia. El análisis con valores SHAP mostró que la predicción se sustenta principalmente en marcadores de gravedad y nivel de intervención en UCI, en concordancia con la literatura.
El estudio demuestra la factibilidad de aplicar técnicas de aprendizaje automático sobre MIMIC-III para apoyar la vigilancia epidemiológica de IAAS en UCI y resalta la importancia de emplear métricas alineadas con la prevalencia real del evento. No obstante, el carácter retrospectivo, unicéntrico y sin validación prospectiva en tiempo real limita la generalización de los modelos, por lo que se requieren estudios adicionales y validaciones externas antes de considerar su integración en la práctica clínica.
Healthcare-associated infections (HAIs) are a critical patient safety issue, particularly in intensive care units (ICUs), where severe clinical conditions, the use of invasive devices and prolonged stays increase morbidity, mortality and hospital costs. Despite surveillance programs, traditional strategies show limitations for the early detection of these events in settings with high workload and large volumes of clinical data. This work develops and evaluates machine learning models for the early prediction of HAIs during ICU stays using the public MIMIC-III v1.4 database, which contains anonymized clinical records from critically ill patients at the Beth Israel Deaconess Medical Center. A cohort of 28 763 ICU stays of adult patients with a length of stay ≥ 48 h and an HAI prevalence close to 3.3% was selected. Labels for CLABSI, CAUTI and PVAP were derived from CDC/NHSN surveillance criteria adapted to MIMIC-III. A set of predictive variables was constructed from demographic, physiological, laboratory, microbiology and procedure information, including the creation of derived tables, data cleaning, plausible-range handling, preprocessing and explicit treatment of class imbalance. Baseline models and classifiers based on Random Forest and XGBoost were implemented, combined with class weighting, undersampling (NearMiss) and oversampling (SMOTE). Performance was assessed using stratified cross-validation and a test set with original prevalence, relying on metrics tailored to rare events (AUPRC, F2, recall), in addition to AUROC. The best results were obtained with a weighted baseline XGBoost model, achieving an AUPRC of 0.097 (versus 0.033 for trivial models), an AUROC of 0.752, a recall of 0.593 and an F2 of 0.265, evidencing a real, albeit moderate, ability to prioritize patients at higher risk of HAIs in a low-prevalence scenario. SHAP value analysis showed that predictions are mainly driven by markers of severity and level of intervention in the ICU, in line with the clinical literature. The study demonstrates the feasibility of applying machine learning techniques on MIMIC-III to support epidemiological surveillance of HAIs in ICUs and underscores the importance of using metrics aligned with the true event prevalence. Nevertheless, its retrospective, single-center nature and the absence of real-time prospective validation limit the generalizability of the models, highlighting the need for additional studies and external validation before considering their integration into clinical practice.
Healthcare-associated infections (HAIs) are a critical patient safety issue, particularly in intensive care units (ICUs), where severe clinical conditions, the use of invasive devices and prolonged stays increase morbidity, mortality and hospital costs. Despite surveillance programs, traditional strategies show limitations for the early detection of these events in settings with high workload and large volumes of clinical data. This work develops and evaluates machine learning models for the early prediction of HAIs during ICU stays using the public MIMIC-III v1.4 database, which contains anonymized clinical records from critically ill patients at the Beth Israel Deaconess Medical Center. A cohort of 28 763 ICU stays of adult patients with a length of stay ≥ 48 h and an HAI prevalence close to 3.3% was selected. Labels for CLABSI, CAUTI and PVAP were derived from CDC/NHSN surveillance criteria adapted to MIMIC-III. A set of predictive variables was constructed from demographic, physiological, laboratory, microbiology and procedure information, including the creation of derived tables, data cleaning, plausible-range handling, preprocessing and explicit treatment of class imbalance. Baseline models and classifiers based on Random Forest and XGBoost were implemented, combined with class weighting, undersampling (NearMiss) and oversampling (SMOTE). Performance was assessed using stratified cross-validation and a test set with original prevalence, relying on metrics tailored to rare events (AUPRC, F2, recall), in addition to AUROC. The best results were obtained with a weighted baseline XGBoost model, achieving an AUPRC of 0.097 (versus 0.033 for trivial models), an AUROC of 0.752, a recall of 0.593 and an F2 of 0.265, evidencing a real, albeit moderate, ability to prioritize patients at higher risk of HAIs in a low-prevalence scenario. SHAP value analysis showed that predictions are mainly driven by markers of severity and level of intervention in the ICU, in line with the clinical literature. The study demonstrates the feasibility of applying machine learning techniques on MIMIC-III to support epidemiological surveillance of HAIs in ICUs and underscores the importance of using metrics aligned with the true event prevalence. Nevertheless, its retrospective, single-center nature and the absence of real-time prospective validation limit the generalizability of the models, highlighting the need for additional studies and external validation before considering their integration into clinical practice.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Biomédico/a.
Keywords
Atención al paciente, Cuidados intensivos, Aprendizaje