PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
February 22, 2026Journal of Mechanics in Medicine and Biology0 citations

Analysis of Mortality Risk Factors and Ensemble Prediction Modeling for DKD Patients in ICU

View Full Paper
YLYanting LinYDYiwen DengHLHaolin Liu

Key Result

An ensemble model combining SVM, RF, and others predicted in-hospital mortality in ICU DKD patients with AUC of 0.868, improving over single models by ~1%.

Key Points

  • This analysis aims to identify risk factors linked to in-hospital mortality for DKD patients in ICU and to create a predictive model.
  • Utilized the MIMIC-IV database to select 434 DKD patient records
  • Employed feature engineering for data preprocessing
  • Performed univariate and multivariate analyses to find significant predictors
  • Evaluated eight individual machine learning models for performance
  • Developed a dynamic weighted ensemble framework for prediction.
  • SVM + RF combination achieved an AUC of 0.865, improving RF's performance by 0.9 percentage points
  • Final ensemble model reached an AUC of 0.8676, a 0.26 percentage point improvement over pairwise combinations
  • Key predictors of mortality include SOFA scores, sodium levels, creatinine levels, and potassium levels.

Structured PICO

Does a dynamic weighted ensemble machine learning framework improve the prediction of in-hospital mortality in ICU patients with diabetic kidney disease compared to single models?

P
Population
434 patients with diabetic kidney disease (DKD) admitted to the intensive care unit (ICU) from the MIMIC-IV database
I
Intervention
Dynamic weighted ensemble machine learning framework
C
Comparator
Single machine learning models (Decision Tree, Random Forest, K-Nearest Neighbors, Support Vector Machine, Adaptive Boosting, eXtreme Gradient Boosting, Naïve Bayes, Multi-Layer Perceptron) and pairwise combinations
O
Outcome
In-hospital mortalityhard clinical

A dynamic weighted ensemble machine learning framework incorporating multiple models significantly improves the prediction of in-hospital mortality risk for ICU patients with diabetic kidney disease.

Abstract

This study, based on the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, aimed to investigate the risk factors associated with in-hospital mortality among patients with diabetic kidney disease (DKD) admitted to the intensive care unit (ICU), and to develop a high-precision predictive model. A total of 434 DKD patient records were meticulously selected through strict inclusion criteria. A filter-based feature engineering strategy was employed. Initially, multi-source heterogeneous data tables were integrated using Structured Query Language (SQL) queries in combination with Python for preprocessing. Missing values in numerical variables were imputed using Bayesian regression or median substitution, and all continuous variables were standardized. Subsequently, key prognostic features were identified through univariate statistical testing (t-test or 𝜒 2 test) followed by multivariate logistic regression analysis. The significant predictors of mortality included demographic characteristics (e. g. , age), laboratory parameters e. g. , the Minimum White Blood Cell count (WBCₘin), the Maximum Blood Urea Nitrogen (BUNₘax), the Mean Sodium (Naₘean), the Sequential Organ Failure Assessment (SOFA) score, clinical interventions (e. g. , mechanical ventilation, vasopressor use), and comorbidities e. g. , hypertension, chronic obstructive pulmonary disease (COPD). In terms of model construction strategy, this study first evaluated the performance of eight individual machine-learning models, including Decision Tree (DT), Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), Naïve Bayes (NB), and Multi-Layer Perceptron (MLP). An innovative pairwise ensemble strategy was applied, resulting in 28 unique model combinations. Several combinations demonstrated significantly improved performance compared with single models. Notably, the SVM + RF combination achieved the highest performance with an Area Under the Curve (AUC) of 0. 865, representing a 0. 9 percentage point improvement over the best-performing single model, RF (AUC = 0. 856). Inspired by these findings, a comprehensive ensemble framework incorporating all eight models was subsequently developed. The final ensemble prediction framework adopted a dynamic weighting strategy based on validation-set AUCs, with hyper-parameter optimization conducted using Optuna via Tree-structured Parzen Estimator (TPE) sampling over 100 iterations. The resulting model weights were as follows: SVM (0. 86), XGBoost (0. 58), RF (0. 69), KNN (0. 52), NB (0. 40), MLP (0. 40), AdaBoost (0. 01), and DT (0. 002). After weighted probability fusion, the ensemble model achieved an AUC of 0. 8676 on the test set, representing a 0. 26 percentage point improvement over the best-performing pairwise combination (SVM + RF). Feature importance analysis identified the average SOFA score (sofaₛcoreₐvg), the mean sodium level (Sodiumₘean), the minimum creatinine level (Creatinineₘin), and the mean potassium level (Potassiumₘean) as the most critical features in predicting in-hospital mortality risk for DKD patients. Additionally, higher SOFA scores, lower WBCₘin, and BUNₘax were associated with an increased risk of mortality, consistent with the results from multivariable logistic regression analysis. This study effectively utilized filter-based feature engineering to identify key prognostic indicators and, inspired by pairwise model-combination results, developed a dynamic weighted ensemble framework that significantly enhanced the identification of high-risk patients.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Lin et al. (2026) studied this question. An ensemble model combining SVM, RF, and others predicted in-hospital mortality in ICU DKD patients with AUC of 0.868, improving over single models by ~1%.

synapsesocial.com/papers/699a9d50482488d673cd32d1https://doi.org/10.1142/s0219519426400348
Ask AI
Helpful
Bookmark
Share
View Full Paper