Resource-Conscious Modeling for Next-Day Discharge Prediction Using Clinical Notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Healthcare Operations

Background:

• Accurate and timely discharge prediction is essential for efficient surgical unit management.
• Predicting patient discharge facilitates resource allocation and reduces patient length of stay.
• Imbalanced datasets, common in healthcare, pose challenges for predictive modeling.

Purpose of the Study:

• To compare the effectiveness of different text analysis models for predicting next-day patient discharge.
• To evaluate traditional methods like TF-IDF against modern approaches such as sentence embeddings and fine-tuned small language models (SLMs).

Main Methods:

• Utilized a dataset of 3,928 postoperative patient notes.
• Compared Term Frequency-Inverse Document Frequency (TF-IDF) with LightGBM, sentence embeddings, and LoRA-fine-tuned SLMs.
• Assessed model performance using F1-score and Area Under the Curve (AUC) metrics on an imbalanced dataset.

Main Results:

• TF-IDF models combined with LightGBM achieved the highest performance (F1 = 0.47, AUC = 0.80).
• Sentence embeddings and SLMs demonstrated limited sensitivity for this specific task.
• LoRA fine-tuning offered only marginal improvements in recall for the transformer-based models.

Conclusions:

• Lightweight and interpretable text models, such as TF-IDF, are more effective than complex transformer-based models for imbalanced next-day discharge prediction.
• Simpler models offer a practical and efficient solution for surgical unit operations.
• Further research could explore hybrid approaches or feature engineering for improved performance.