Retrieving Evidence from EHRs with LLMs: Possibilities and Challenges

Overview

• This paper explores using large language models (LLMs) to efficiently retrieve and summarize relevant information from unstructured Electronic Health Records (EHRs) to aid radiologists' diagnoses.
• The key idea is to task an LLM with inferring whether a patient has or is at risk of a particular condition based on their EHR notes, and then ask the model to summarize the supporting evidence.
• The researchers evaluate this approach against a pre-LLM information retrieval baseline and find that the LLM-based method produces outputs that are consistently preferred by experts.
• They also propose and validate a method to use an LLM to evaluate other LLM outputs for this task, allowing them to scale up the evaluation process.

Plain English Explanation

Electronic Health Records (EHRs) often contain valuable information that could help radiologists make more informed diagnoses. However, the huge volume of notes associated with each patient, coupled with the time constraints radiologists face, makes it practically impossible for them to manually identify all the relevant evidence.

The researchers in this study propose using large language models (LLMs) as a way to efficiently retrieve and summarize the unstructured information in EHRs that could be relevant to a given medical query. The idea is to task the LLM with determining whether a patient has or is at risk of a particular condition, and then ask it to provide a summary of the evidence from the patient's notes that supports that assessment.

When the researchers had experts evaluate the outputs of this LLM-based approach, they found that the experts consistently preferred these outputs over the results of a traditional information retrieval system. This suggests that LLMs can be effective interfaces for accessing the wealth of information contained in EHRs.

Since manually evaluating all the LLM outputs would be extremely time-consuming, the researchers also developed a way to use an LLM to automatically evaluate other LLM outputs for this task. This allows them to scale up the evaluation process and get a better sense of the overall performance of their approach.

Overall, this research indicates the promise of using LLMs to help radiologists more easily access and make use of the valuable information hidden in EHR notes. However, it also highlights the challenge of dealing with LLM outputs that may include fabricated or inaccurate information, known as "hallucinations." The researchers show that in this particular setting, the LLM's own confidence in its outputs can be a useful signal for identifying faithful summaries and limiting confabulations.

Technical Explanation

The key innovation in this work is the use of a zero-shot strategy that tasks an LLM with inferring whether a patient has or is at risk of a particular condition based on their EHR notes, and then asking the model to summarize the supporting evidence.

To evaluate this approach, the researchers compared it to a pre-LLM information retrieval baseline. They found that under expert evaluation, the LLM-based outputs were consistently preferred over the baseline.

Manual evaluation is expensive, so the researchers also proposed and validated a method using an LLM to evaluate (other) LLM outputs for this task. This allowed them to scale up the evaluation process and get a more comprehensive understanding of the performance of their approach.

The findings indicate the promise of LLMs as interfaces to EHRs, but also highlight the challenge of dealing with hallucinations - outputs that contain fabricated or inaccurate information. However, the researchers show that in this setting, the model's own confidence in its outputs strongly correlates with the faithfulness of the summaries, providing a practical way to limit confabulations.

Critical Analysis

The researchers acknowledge several limitations and areas for further research. First, they note that their evaluation was limited to a small set of conditions, and more comprehensive testing is needed to fully understand the capabilities and limitations of their approach.

Additionally, while they demonstrate that model confidence can be used to identify faithful summaries, this is not a perfect solution. There may still be cases where the model is overconfident in its incorrect outputs. Further research is needed to develop more robust methods for detecting and mitigating hallucinations in this context.

Another potential issue is the reliance on expert manual evaluation, which is time-consuming and expensive. While the researchers' proposed method for using an LLM to evaluate other LLM outputs is a step in the right direction, more work is needed to fully automate the evaluation process and make it scalable.

Finally, the researchers do not address the potential ethical and privacy concerns that may arise from using LLMs to access and summarize sensitive patient data. As these technologies become more widely adopted, it will be crucial to carefully consider the implications and develop appropriate safeguards.

Conclusion

This research demonstrates the potential of using large language models to efficiently extract and summarize relevant information from unstructured Electronic Health Records to support radiologists' diagnoses. The LLM-based approach outperformed a traditional information retrieval baseline, and the researchers developed a novel method to scale up the evaluation process.

However, the study also highlights the challenge of dealing with hallucinations - LLM outputs that contain fabricated or inaccurate information. While the researchers showed that model confidence can be a useful signal for identifying faithful summaries, more work is needed to fully address this issue.

Overall, this research represents an important step forward in leveraging the power of large language models to unlock the valuable information contained in unstructured medical data. As these technologies continue to evolve, it will be crucial to carefully consider the ethical and practical implications to ensure they are deployed in a responsible and beneficial manner.