Development, Validation, and Deployment of a Time-Dependent Machine Learning Model for Predicting One-Year Mortality Risk in Critically Ill Patients with Heart Failure DOI Creative Commons
Jiuyi Wang,

Qingxia Kang,

Shiqi Tian

et al.

Bioengineering, Journal Year: 2025, Volume and Issue: 12(5), P. 511 - 511

Published: May 12, 2025

Background: Heart failure (HF) ranks among the foremost causes of mortality globally, exhibiting particularly high prevalence and significant impact within intensive care units (ICUs). This study sought to develop, validate, deploy a time-dependent machine learning model aimed at predicting one-year all-cause risk in ICU patients diagnosed with HF, thereby facilitating precise prognostic evaluation stratification. Methods: encompassed cohort 8960 HF sourced from Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 3.1). latest version added data 2020 2022 on basis 2.2 (covering 2008 2019); therefore, spanning 2019 (n = 5748) were designated training set, while 3212) reserved test set. The primary endpoint interest was mortality. Least Absolute Shrinkage Selection Operator (LASSO) regression employed select predictive features an initial pool 64 candidate variables (including demographic characteristics, vital signs, comorbidities complications, therapeutic interventions, routine laboratory data, disease severity scores). Four models developed compared: Cox proportional hazards, random survival forest (RSF), hazards deep neural network (DeepSurv), eXtreme Gradient Boosting (XGBoost). Model performance assessed using concordance index (C-index) Brier score, interpretability addressed through SHapley Additive exPlanations (SHAP) Survival (SurvSHAP(t)). Results: revealed rate 46.1% population under investigation. In LASSO effectively identified 24 model. XGBoost exhibited superior performance, as evidenced by C-index 0.772 score 0.161, outperforming (C-index: 0.740, score: 0.175), RSF 0.747, 0.178), DeepSur 0.723, 0.183). Decision curve analysis validated clinical utility across broad spectrum thresholds. Feature importance red cell distribution width-to-albumin ratio (RAR), Charlson Comorbidity Index, Simplified Acute Physiology Score II (SAPS II), III (APS III), age-bilirubin-INR-creatinine (ABIC) top five factors. Consequently, online prediction tool based this has been is publicly accessible. Conclusions: demonstrated robust capability evaluating critically ill patients. offered useful early identification supported timely interventions.

Language: Английский

Impact of the Geriatric Nutritional Risk Index on short-term prognosis of patients with sepsis-related acute kidney injury: analysis using the MIMIC-IV database DOI Creative Commons

Kailun Cai,

Wenchao Mao, Mingkun Yang

et al.

BMC Nephrology, Journal Year: 2025, Volume and Issue: 26(1)

Published: April 23, 2025

Language: Английский

Citations

0
Development, Validation, and Deployment of a Time-Dependent Machine Learning Model for Predicting One-Year Mortality Risk in Critically Ill Patients with Heart Failure DOI Creative Commons
Jiuyi Wang,

Qingxia Kang,

Shiqi Tian

et al.

Bioengineering, Journal Year: 2025, Volume and Issue: 12(5), P. 511 - 511

Published: May 12, 2025

Background: Heart failure (HF) ranks among the foremost causes of mortality globally, exhibiting particularly high prevalence and significant impact within intensive care units (ICUs). This study sought to develop, validate, deploy a time-dependent machine learning model aimed at predicting one-year all-cause risk in ICU patients diagnosed with HF, thereby facilitating precise prognostic evaluation stratification. Methods: encompassed cohort 8960 HF sourced from Medical Information Mart for Intensive Care IV (MIMIC-IV) database (version 3.1). latest version added data 2020 2022 on basis 2.2 (covering 2008 2019); therefore, spanning 2019 (n = 5748) were designated training set, while 3212) reserved test set. The primary endpoint interest was mortality. Least Absolute Shrinkage Selection Operator (LASSO) regression employed select predictive features an initial pool 64 candidate variables (including demographic characteristics, vital signs, comorbidities complications, therapeutic interventions, routine laboratory data, disease severity scores). Four models developed compared: Cox proportional hazards, random survival forest (RSF), hazards deep neural network (DeepSurv), eXtreme Gradient Boosting (XGBoost). Model performance assessed using concordance index (C-index) Brier score, interpretability addressed through SHapley Additive exPlanations (SHAP) Survival (SurvSHAP(t)). Results: revealed rate 46.1% population under investigation. In LASSO effectively identified 24 model. XGBoost exhibited superior performance, as evidenced by C-index 0.772 score 0.161, outperforming (C-index: 0.740, score: 0.175), RSF 0.747, 0.178), DeepSur 0.723, 0.183). Decision curve analysis validated clinical utility across broad spectrum thresholds. Feature importance red cell distribution width-to-albumin ratio (RAR), Charlson Comorbidity Index, Simplified Acute Physiology Score II (SAPS II), III (APS III), age-bilirubin-INR-creatinine (ABIC) top five factors. Consequently, online prediction tool based this has been is publicly accessible. Conclusions: demonstrated robust capability evaluating critically ill patients. offered useful early identification supported timely interventions.

Language: Английский

Citations

0