Machine learning-driven prediction of hospital admissions using gradient boosting and GPT-2.

Abstract

Background: Accurately predicting hospital admissions from the emergency department (ED) is essential for improving patient care and resource allocation. This study aimed to predict hospital admissions by integrating both structured clinical data and unstructured text data using machine learning models.

Methods: Data were obtained from the 2021 National Hospital Ambulatory Medical Care Survey-Emergency Department (NHAMCS-ED), including adult patients aged 18 years and older. Structured data included demographics, visit characteristics, vital signs, and medical history, while unstructured data consisted of free-text chief complaints and injury descriptions. A Gradient Boosting Classifier (GBC) was applied to structured data, while a fine-tuned GPT-2 model processed the unstructured text. A combined model was created by averaging the outputs of both models. Model performance was evaluated using 5-fold cross-validation, assessing accuracy, precision, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC-ROC).

Results: Among the 13,115 patients, 2264 (17.3%) were admitted to the hospital. The combined model outperformed the individual structured and unstructured models, achieving an accuracy of 75.8%, precision of 39.5%, sensitivity of 75.8%, and specificity of 75.8%. In comparison, the structured data model achieved 73.8% accuracy, while the unstructured model reached 64.6%. The combined model had the highest AUC, indicating superior performance.

Conclusions: Combining structured and unstructured data using machine learning significantly improves the prediction of hospital admissions from the ED. This integrated approach can enhance decision-making and optimize ED operations.