Journal Information
Share
Download PDF
More article options
Original article
Available online 4 November 2025
Toward transparent clinical decision support in the ICU: A multi-window, model-agnostic explainability approach for 72-h mortality prediction using the eICU collaborative research database
Hacia un apoyo transparente a la toma de decisiones clínicas en la UCI: un enfoque de explicabilidad multiventana e independiente del modelo para la predicción de la mortalidad a 72 horas utilizando la base de datos de investigación colaborativa eICU
Daniel Gallardo Gómeza,b,1, Sara Díazb,1, Javier Quintero Sosaa,b,1, Claudia Jiménez Vázquezb, Álvaro Ritoré-Hidalgob, Antonio Gutiérrez-Pizarrayaa,c, Miguel Ángel Armengol de la Hozb,
, María Recuerda Núñezd,2, Ángel Estellad,e,f,2
a Health Technology Assessment Area (AETSA), Progress and Health Public Foundation (FPS), Seville, Spain
b Data Science Lab, Progress and Health Public Foundation (FPS), Seville, Spain
c Research Group IBiS CS-017 'Lung Cancer Research, Seville, Spain
d Intensive Care Unit, University Hospital of Jerez, Jerez, Spain
e Department of Medicine, University of Cádiz, Cádiz, INIBiCA, Spain
f Biomedical Research and Innovation Institute of Cádiz (INiBICA) Research Unit, University of Cádiz, Cádiz, Spain
Article information
Abstract
Full Text
Bibliography
Download PDF
Statistics
Figures (3)
Show moreShow less
Tables (5)
Table 1. Constant study sample characteristics.
Tables
Table 2. Time-dependent study sample characteristics.
Tables
Table 3. Balanced accuracy, F1-score and recall for different data windows and models.
Tables
Table 4. Use cases of the clinical decision-support system.
Tables
Table 5. F1-scores obtained for the test dataset and for groups for each data window.
Tables
Show moreShow less
Additional material (1)
Abstract
Objective

This study has two main objectives: (1) to develop a multi-model framework for predicting Intensive Care Unit (ICU) mortality within the first 72 h of admission; and (2) to introduce a novel model-agnostic explainability approach classification that enables variable-level interpretation of predicted probabilities.

Design

Retrospective study using a multi-model machine learning approach, analyzing data across multiple time windows and incorporating demographic, clinical, and biochemical variables.

Setting

ICU mortality.

Patients or participants

Patients included in the eICU database over 16 years old who have been admitted to ICUs in 2014 and 2015 with available data within the first 72 h after ICU admission. A total of 106,449 patients were included in the analyses.

Interventions

No clinical interventions were applied; this was a retrospective analysis for predictive model development and evaluation.

Main variables of interest

Demographic, clinical, and biochemical variables collected across multiple time windows.

Results

A total of 106,449 were included (mean age 62.6 years, 46% women), with an overall 72-h mortality of 4.8%. Random Forest models achieved one of the best results in terms of predictive performance metrics, with F1-scores of 0.93, 95% CI 0.93 to 0.94; 0.92, 95% CI 0.92 to 0.93 and 0.83, 95% CI 0.83 to 0.85 across the three temporal data windows. Due to these metrics, the ability to predict deaths, and the biological plausibility of the predictions, Random Forest models were selected from all those studied.

Conclusions

The proposed multi-model approach significantly improves 72-h ICU mortality prediction. Moreover, we outline a model-agnostic strategy for variable-level interpretation of predicted probabilities, which may facilitate transparency and support future applications in clinical decision support.

Keywords:
Intensive Care Unit
Mortality
Database
Machine learning
Clinical decision support system
Resumen
Objetivo

Este estudio tiene dos objetivos principales: (1) desarrollar un multimodelo para predecir la mortalidad en la Unidad de Cuidados Intensivos (UCI) dentro de las primeras 72 horas de ingreso; e (2) introducir una novedosa metodología aplicable a cualquier modelo de clasificación que permita la interpretación a nivel variable de las probabilidades predichas.

Diseño

Estudio retrospectivo utilizando un enfoque de aprendizaje automático multimodelo, analizando los datos a través de múltiples ventanas temporales e incorporando variables demográficas, clínicas y bioquímicas.

Ámbito

mortalidad en UCI.

Pacientes o participantes

Pacientes incluidos en la base de datos eICU mayores de 16 años ingresados en UCI en 2014 y 2015 con datos disponibles en las primeras 72 horas tras el ingreso en UCI, lo que significa que se incluyeron en los análisis un total de 106.449 pacientes.

Intervenciones

No se aplicaron intervenciones clínicas; se trató de un análisis retrospectivo para el desarrollo y la evaluación de modelos predictivos.

Variables de interés principals

Variables demográficas, clínicas y bioquímicas recogidas a lo largo de múltiples ventanas temporales.

Resultados

Se incluyeron 106,449 pacientes (edad media 62,6 años; 46% mujeres), con una mortalidad global a 72 horas del 4,8%. Los modelos Random Forest consiguieron uno de los mejores resultados en cuanto a métricas de rendimiento predictivo, con F1-score de 0.93, IC del 95 %: 0.93 a 0.94; 0.92, IC del 95 %: 0.92 a 0.93 y 0.83, IC del 95 %: 0.83 a 0.85 en las tres ventanas de datos temporales. Debido a estas métricas, la capacidad de predecir muertes y la plausibilidad biológica de las predicciones, se seleccionaron los modelos de Random Forest de entre todos los estudiados.

Conclusiones

El enfoque multimodelo propuesto mejora significativamente la predicción de la mortalidad en la UCI a las 72 horas. Asimismo, presentamos una estrategia agnóstica al modelo para la interpretación a nivel de variable de las probabilidades predichas, que puede contribuir a la transparencia y servir de base para futuras aplicaciones en sistemas de apoyo a la decisión clínica.

Palabras clave:
Unidad de cuidados intensivos
Mortalidad
Base de datos
Aprendizaje automático
Sistema de apoyo a la decisión clínica

Article

These are the options to access the full texts of the publication Medicina Intensiva (English Edition)
Member
If you are a member of the Sociedad Española de Medicina Intensiva, Crítica y Unidades Coronarias::
Go to the members area of the website of the SEMICYUC (www.semicyuc.org )and click the link to the magazine.
Subscriber
Subscriber

If you already have your login data, please click here .

If you have forgotten your password you can you can recover it by clicking here and selecting the option “I have forgotten my password”
Subscribe
Subscribe to

Medicina Intensiva (English Edition)

Purchase
Purchase article

Purchasing article the PDF version will be downloaded

Purchase now
Contact
Phone for subscriptions and reporting of errors
From Monday to Friday from 9 a.m. to 6 p.m. (GMT + 1) except for the months of July and August which will be from 9 a.m. to 3 p.m.
Calls from Spain
932 415 960
Calls from outside Spain
+34 932 415 960
Email
Idiomas
Medicina Intensiva (English Edition)
Article options
Tools
Supplemental materials