Scholarlink

Risk stratification of Intensive Care Units (ICUs) is vital in order to make timely interventions and minimize mortality rates. We designed a two-phase prediction model to predict in-hospital short-term (24–48 h) and long-term (30+90 days) mortality with the help of MIMIC-IV v3.1 data, based on 1.76 million inpatient records. With feature engineering, which was our rigorous process, we engineered heterogeneous features, i.e. demographics, vital signs, lab values and ICD codes- to an elegant save 33-feature set. We compared XGBoost and LightGBM models with stable settings, tuned models with Optuna, and the long-term XGBoost was the only model that improved and helped to boost the overall performance. The combined model demonstrated solid results that were free of feature leakage and renal failure and infectious disease were identified as significant mortality predictors. This predictive and interpretable model provides clinicians working in the ICU with data-driven insights and enables timely interventions as well as better patient outcomes due to enhanced decision-making.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MIMIC-IV ICU Risk Stratification: A Dual-Stage Predictive Model for Short and Long-Term Patient Outcome Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Neetesh Jatav,
  • Vibha Tiwari,
  • Aryan Singh

摘要

Risk stratification of Intensive Care Units (ICUs) is vital in order to make timely interventions and minimize mortality rates. We designed a two-phase prediction model to predict in-hospital short-term (24–48 h) and long-term (30+90 days) mortality with the help of MIMIC-IV v3.1 data, based on 1.76 million inpatient records. With feature engineering, which was our rigorous process, we engineered heterogeneous features, i.e. demographics, vital signs, lab values and ICD codes- to an elegant save 33-feature set. We compared XGBoost and LightGBM models with stable settings, tuned models with Optuna, and the long-term XGBoost was the only model that improved and helped to boost the overall performance. The combined model demonstrated solid results that were free of feature leakage and renal failure and infectious disease were identified as significant mortality predictors. This predictive and interpretable model provides clinicians working in the ICU with data-driven insights and enables timely interventions as well as better patient outcomes due to enhanced decision-making.