SPeC: A Soft Prompt-Based Calibration on Performance Variability of Large Language Model in Clinical Notes Summarization

Yu-Neng Chuang, Ruixiang Tang, Xiaoqian Jiang, Xia Hu

TL;DR

This paper introduces SPeC, a novel framework that uses "soft prompts" to stabilize the performance of large language models (LLMs) when summarizing clinical notes. By calibrating LLM outputs, SPeC significantly reduces variability without sacrificing summarization quality, making LLMs more reliable for healthcare professionals.

What Problem Does This Paper Address?

Electronic Health Records (EHRs) contain a wealth of patient information, but manually summarizing clinical notes is time-consuming and prone to errors. While Large Language Models (LLMs) show great promise for automating this task, especially when guided by carefully crafted "prompts" (instructions), a significant challenge arises: prompt engineering is difficult, and even minor variations in prompts can lead to highly inconsistent and unpredictable summary outputs. This variability makes it hard for non-experts to trust and effectively use LLMs for critical tasks like clinical note summarization, potentially leading to overlooked information or errors in patient care.

Methods

The core of this paper is the Soft Prompt-Based Calibration (SPeC) framework. Unlike traditional "discrete" text prompts that have fixed meanings, SPeC utilizes soft prompts. These are learned, flexible vectors in the LLM's embedding space that don't have pre-defined meanings.

Here's how it works:

  1. LLM-Generated Prompts: The process starts by using LLMs (like ChatGPT) to automatically generate a set of diverse text prompts for clinical note summarization. This simulates how non-experts might interact with LLMs.
  2. Soft Prompt Encoder: A novel, model-agnostic soft prompt encoder is introduced. This encoder takes the LLM-generated text prompts and maps them into soft prompt vectors. Crucially, this encoder is trained in a zero-shot manner, meaning it doesn't require specific summarization examples for training.
  3. INFOSUM Loss: To train the soft prompt encoder and reduce variability, a new loss function called INFOSUM Loss is proposed. This loss aims to minimize the divergence between the embedding distributions of the original clinical notes and the summarized outputs, effectively making the LLM's output more consistent regardless of the input prompt's subtle variations.
  4. Calibration: During inference, the learned soft prompt vectors are combined with the LLM-generated text prompts to guide the LLM, leading to more stable and reliable summarization.

The key innovation is the use of soft prompts to act as a calibrator, smoothing out the performance fluctuations caused by the inherent variability of LLM-generated discrete prompts.

Datasets

The primary dataset used in this research is MIMIC-CXR. This is a large dataset containing chest radiographs and their corresponding radiology reports. For this study, the "findings" section of the radiology reports served as the input clinical notes, and the "impressions" section was used as the ground-truth summary. The authors complied with data access requirements and completed necessary training.

Key Insights and Ideas

  • Prompt Variability is a Real Problem: The paper empirically demonstrates that even minor changes in LLM-generated prompts can lead to significant differences in summary quality and content, highlighting a critical bottleneck for practical LLM deployment in sensitive domains.
  • Soft Prompts as Calibrators: The core idea that soft prompts can act as a "calibrator" to stabilize LLM outputs is a powerful concept. They offer a way to inject learned, consistent guidance without needing to retrain the entire LLM.
  • Model-Agnostic Approach: SPeC is designed to be model-agnostic, meaning it can be applied to various LLMs without requiring specific architectural changes, increasing its generalizability.
  • Zero-Shot Learning for Prompts: The soft prompt encoder being trained in a zero-shot manner is significant. It means SPeC can be applied without needing large amounts of labeled summarization data for each new LLM or task, making it more efficient.
  • In-Distribution Prompts Matter: The study found that using "in-distribution" soft prompt tokens (those that align with the domain's language) leads to better performance and variability reduction compared to random or "out-of-distribution" tokens.

Key Results

  • Significant Variability Reduction: SPeC demonstrated a substantial reduction in performance variability. For instance, it reduced variability by up to 43.1% on the Flan-T5 model for ROUGE-L scores.
  • Improved Performance: SPeC not only reduced variability but also, in many cases, slightly improved the overall summarization performance (measured by ROUGE F1-scores) compared to using LLM-generated prompts alone.
  • Consistent Across LLMs: The benefits of SPeC were observed across three different transformer-based LLMs: Flan-T5, BART, and Pegasus-xsum, underscoring its model-agnostic nature.
  • Case Study Validation: A qualitative case study showed that without SPeC, certain prompts could lead to incorrect or incomplete summaries, while SPeC ensured more stable and accurate outputs.

Limitations and Future Work

  • Domain Specificity: The current work focuses on radiology reports, which have a relatively standardized format. Applying SPeC to more complex medical reports with varied terminology and abbreviations might require further adaptation.
  • Privacy Concerns: Summarizing medical reports raises privacy issues. While the authors state they did not input patient data directly into OpenAI platforms for prompt generation, future work needs to ensure robust patient privacy and data security compliance, especially with more sensitive medical data.
  • Human Oversight: The authors emphasize that AI-generated summaries should always be double-verified by trained doctors. Relying solely on AI could lead to overlooked critical findings.
  • Future Directions: The authors plan to extend SPeC to more complex clinical note summarization tasks beyond radiology reports. They also aim to explore more sophisticated methods for generating and selecting soft prompt tokens.

Summary updated

Abstract

Electronic health records (EHRs) store an extensive array of patient information, encompassing medical histories, diagnoses, treatments, and test outcomes. These records are crucial for enabling healthcare providers to make well-informed decisions regarding patient care. Summarizing clinical notes further assists healthcare professionals in pinpointing potential health risks and making better-informed decisions. This process contributes to reducing errors and enhancing patient outcomes by ensuring providers have access to the most pertinent and current patient data. Recent research has shown that incorporating instruction prompts with large language models (LLMs) substantially boosts the efficacy of summarization tasks. However, we show that this approach also leads to increased performance variance, resulting in significantly distinct summaries even when instruction prompts share similar meanings. To tackle this challenge, we introduce a model-agnostic Soft Prompt-BasedCalibration (SPeC) pipeline that employs soft prompts to lower variance while preserving the advantages of prompt-based summarization. Experimental findings on multiple clinical note tasks and LLMs indicate that our method not only bolsters performance but also effectively regulates variance across different LLMs, providing a more consistent and reliable approach to summarizing critical medical information.
BibTeX
@article{chuang2024spec,
  author = {Yu-Neng Chuang and Ruixiang Tang and Xiaoqian Jiang and Xia Hu},
  title = {{SPeC: A Soft Prompt-Based Calibration on Performance Variability of Large Language Model in Clinical Notes Summarization}},
  journal = {J. Biomed. Informatics},
  year = {2024},
  doi = {10.1016/j.jbi.2024.104606},
}
Figure