Aspect-Oriented Summarization for Psychiatric Short-Term Readmission Prediction.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Large language models (LLMs) show promise for processing lengthy documents without task-specific data.
• However, their zero-shot performance on complex tasks remains suboptimal.
• Summarization can reduce information loss for complex document analysis.

Purpose of the Study:

• To develop a method for processing LLM summaries to capture diverse information signals.
• To improve supervised training of transformer models using integrated summary signals.
• To enhance prediction performance for complex patient outcome tasks.

Main Methods:

• Generating LLM summaries using aspect-oriented prompts.
• Measuring information differences between aspect-oriented summaries.
• Integrating signals from multiple summaries for supervised training.
• Validating the approach on 30-day psychiatric readmission prediction.

Main Results:

• Hypothesized that aspect-oriented summaries contain distinct information signals.
• Developed methods to measure and integrate these signals.
• Demonstrated increased prediction performance on a real-world clinical task.
• Showcased improved patient outcome prediction using integrated summary data.

Conclusions:

• Integrating information from multiple aspect-oriented LLM summaries enhances supervised learning.
• This approach effectively addresses information loss in document summarization.
• The method shows significant potential for improving complex predictive tasks in healthcare.
• Future work can explore broader applications in clinical decision support.