ICU Length of Stay Prediction for Patients with Diabetes Using Machine Learning and Clinical Notes.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Diabetes mellitus is a chronic condition associated with increased healthcare utilization and costs.
• Accurate prediction of intensive care unit (ICU) length of stay (LOS) is essential for effective hospital resource allocation and patient management.
• Diabetic patients often experience complex health issues, necessitating precise LOS estimations in critical care settings.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting ICU length of stay (LOS) in diabetic patients.
• To integrate unstructured clinical notes with structured electronic health record data for improved predictive accuracy.
• To compare the performance of different machine learning approaches for both regression (predicting days) and classification (short vs. long stays).

Main Methods:

• Utilized a dataset of diabetic patients admitted to the ICU, incorporating structured data (demographics, diagnoses, lab results, ICU events) and unstructured clinical notes.
• Employed natural language processing techniques, including Doc2Vec word embeddings and TF-IDF text encoding, for feature extraction from clinical notes.
• Developed and compared machine learning models, including neural networks and logistic regression, for predicting ICU LOS.
• Evaluated model performance using metrics such as R-squared, mean absolute error (MAE), and accuracy.

Main Results:

• A neural network model incorporating Doc2Vec word embeddings achieved the best performance for predicting short ICU stays (R2 = 0.3626, MAE = 1.54 days).
• Logistic regression utilizing TF-IDF text encoding demonstrated superior performance for classifying ICU stays as long (≥10 days) or short (<10 days), achieving an accuracy of 0.875.
• The integration of unstructured clinical notes alongside structured data significantly enhanced the predictive capabilities of the developed models.

Conclusions:

• Combining structured clinical data with unstructured information from clinical notes using machine learning offers a powerful approach for early and accurate prediction of ICU LOS in diabetic patients.
• These predictive models can support clinical decision-making, facilitate proactive patient management, and optimize hospital resource allocation.
• Further research into advanced NLP and machine learning techniques holds promise for refining ICU LOS predictions and improving patient outcomes in critical care settings.