Early Intensive Care Unit Length of Stay Prediction on MIMIC-IV: A Dual Approach With Clinical Features and Textual...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Clinical data analysis

Background:

• Accurate prediction of intensive care unit (ICU) length of stay (LOS) within 24 hours of admission is crucial for optimizing hospital resource management, including bed allocation, staffing, and patient care planning.
• Traditional methods rely on structured electronic health record (EHR) data, but recent advancements in transformer-based language models offer the potential to leverage unstructured clinical notes for improved predictive accuracy.

Purpose of the Study:

• To compare the effectiveness of machine learning models using structured EHR data versus transformer-based language models using unstructured clinical notes for early prediction of ICU length of stay (LOS).
• To evaluate both binary classification (short vs. long stay) and regression tasks for predicting exact LOS.

Main Methods:

• Two parallel pipelines were developed using the MIMIC-IV dataset: a structured pipeline employing conventional machine learning models (logistic regression, random forest, XGBoost, SVM) on day-one EHR features with ICD-derived embeddings, and an unstructured pipeline fine-tuning transformer models (ClinicalBERT, Bio+ClinicalBERT, BlueBERT) on discharge notes.
• Model performance was assessed using Area Under the Receiver Operating Characteristic curve (AUROC) for binary classification and other relevant metrics for regression.

Main Results:

• The XGBoost model utilizing ICD embeddings achieved the highest performance for binary classification with an AUROC of 0.805 and minimal training time.
• XGBoost without ICD embeddings demonstrated strong performance (AUROC = 0.732), serving as a baseline that does not rely on potentially delayed coding information.
• Transformer models showed comparable performance (AUROC = 0.766) but required significantly more computational resources.

Conclusions:

• Both structured and unstructured data modeling approaches provide valuable early indicators for ICU length of stay (LOS) prediction.
• The choice between structured (e.g., XGBoost with ICD embeddings) and unstructured (e.g., transformers) approaches involves a trade-off between predictive performance, computational efficiency, and ease of integration into existing clinical workflows.