Improving Hospital Length of Stay Prediction through Heterogeneous Data Integration from MIMIC-III Records.

Area of Science:

• * Clinical informatics and predictive analytics.
• * Health services research and operational efficiency.

Background:

• * Accurate prediction of hospital length of stay (LoS) is essential for optimizing hospital operations and patient care.
• * Existing models often rely on limited data types, potentially hindering predictive accuracy.

Purpose of the Study:

• * To comprehensively evaluate machine learning models for LoS classification using multimodal data.
• * To assess the impact of integrating structured clinical data, symptoms, physiological signals, and clinical notes.
• * To compare the performance of various models and feature reduction techniques.

Main Methods:

• * Construction of seven data configurations using structured features (Z), symptoms (S), physiological signals (F), and textual notes (E).
• * Application of five machine learning models: Artificial Neural Networks (ANN), XGBoost, Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM).
• * Implementation of feature selection and Synthetic Minority Over-sampling Technique (SMOTE) for model optimization.

Main Results:

• * The structured feature set (Z) alone provided strong baseline performance.
• * Integrating additional data types (S, F, E) consistently enhanced predictive performance, with the ZSEF configuration achieving top results.
• * XGBoost and ANN models demonstrated superior generalizability across different data configurations.
• * SMOTE improved performance in disease-specific cohorts, notably for lung cancer patients.

Conclusions:

• * Multimodal data integration significantly advances the accuracy of hospital length of stay prediction.
• * Feature reduction techniques are effective in optimizing predictive models.
• * Machine learning models, particularly XGBoost and ANN, show promise for improving healthcare resource management and patient flow.