Enhancing the prediction of hospital discharge disposition with extraction-based language model classification

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Healthcare Predictive Analytics

Background:

• Effective care transition planning requires early identification of patients discharged to skilled nursing facilities (SNFs).
• Predictive clinical information is often fragmented across lengthy admission history and physical (H&P) notes.
• Existing language models face challenges with long documents, noisy data, and lack of transparency.

Purpose of the Study:

• To develop and evaluate extraction-based language model classification (ELC) for predicting SNF discharges.
• To assess if AI Risk Snapshots derived from ELC improve language model performance compared to raw H&P text.
• To mitigate limitations of language models in processing extensive clinical documentation.

Main Methods:

• Developed ELC to distill H&Ps into structured data and concise AI Risk Snapshots.
• Retrospectively compared nine language models using raw H&P text, truncated notes, Structured Extracted Data, and AI Risk Snapshots.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC).

Main Results:

• ELC significantly reduced input data length (AI Risk Snapshot median 141 tokens vs. raw H&P median 2,120 tokens).
• Average AUROC and AUPRC improved across models when using ELC-derived predictors.
• Bio+Clinical BERT fine-tuned on AI Risk Snapshots achieved the highest AUROC of 0.851.

Conclusions:

• ELC effectively structures and summarizes H&P notes, overcoming language model token length constraints.
• AI Risk Snapshots enhance prediction performance and interpretability, aligning with clinical assessments.
• ELC represents a promising approach to improve SNF discharge prediction and facilitate care transitions.