Medical Record Summarization
Medical record summarization is the task of using language processing to quickly represent the content of a medical record. Medical Record Review (MRR) drives the need for this, when medical records are manually analyzed for the purposes of primary care, HEDIS, risk adjustment, prior authorization, or disability, and can range to as high as 50,000 pages in length. Summarization persists today as the #1 largest use of LLM NLP technology in healthcare. Quickly understanding content can give a reviewer a key picture of who and what a record is about is before delving in for more detail.
Tenasol, as a regular part of its business performs medical record summarization with multiple strategies described below, tailored to the client. Note the following:
There many methods to summarize a medical record or group of records: We will cover techniques, parameters, and output formats in this document.
Medical record summarization will always lose data. Inherently by summarizing data detailed information is removed. The exception of this is duplicate information removal.
Medical record summarization is less effective for OCR data. OCR data, directly produced from images that often include tables, images, handwriting, variation in text formatting/positioning, and forms often do not contain full sentence structure, and are inherently more difficult for summarization techniques to evaluate.
Inputs for Medical Record Summarization
Medical data can come in HL7 FHIR, HL7 CDA, HL7 ADT, or image unstructured format (PDF/Image). Depending on the method of summarization these types may need to be preprocessed before they can but processed through a summarization model.
This is especially true if the model is not a LLM/GPT/Transformer model. Furthermore, if seeking a data-based summary, LLM transformers will underperform as they are less capable of quantifying data or expressing confidence scores associated with findings.
Approaches for Medical Record Summarization
There are generally 3 approaches
“Extraction“-based medical record summarization: Statistical summarization methods have been around long for a long time. “Extraction“ means the most important sentences are extracted and presented to the user, rather than a generated summary.
These approaches seek to determine the most valuable sentences by assessing the content of each sentence. Sentences are given more priority if they have rare phrases that have high similarity to many other sentences in a document.
The proliferation of LLM systems have largely rendered these redundant. These systems cannot generate context specific summarizations, but they can summarize to a user-specified sentence length.
“Extraction“-based visual medical record summarization: Data extracted from a medical record using NLP approaches (as well as structured data extraction when applicable) may be used to generate statistical reports. See the example below. This may be used to:
Create map visualizations of locations and their importance
Build visual representations of timelines
Build visualizations associated with data sources used
Depict quantified machine learning confidences of models
Generate statistics
Tenasol visual data summary of medical record
“Abstraction“-based OCR LLM medical record summarization: These systems simply use an LLM (sliding window text encoder and decoder transformer) to create an “abstraction“ (as opposed to extraction) summary based on text that has already been extracted from an image using OCR. These systems further permit the user to specify both context and length of the output. Tenasol performs OCR LLM summarization as a regular part of its services.
It is worth noting that there are varying levels of quality of LLM medical record summarization as a function of compute that a party decides to use.
“Abstraction“-based non-OCR LLM medical record summarization: This is an uncommon approach that uses an image to directly output LLM GPT response using a creative neural network architecture. Stated another way, no OCR is required.
The most popular model at present is named Donut. These models uniquely offer lower costs by bypassing OCR fees, at the cost of lower accuracy. However, they are capable of picking up some context clues that OCR-based LLM techniques may miss.
Hurdles with LLM Medical Record Summarization
Given that the state-of-the-art summarization techniques for medical record summarization make use of LLM technology, there are a few issues that one must note:
Input Limit and Output Limit: When summarizing a record with an LLM, there are input and output size limits. This means that there is a limit of the amount of data that can be input (record data + query) which can be made when performing a summarization on an input text. On the output, if there is more text than the LLM can output, information will be excluded from the result.
Batching Problem: Medical records can reach lengths that are well beyond the upper limits permitted by an LLM system. This means that an approach must be made to break down a record, to make a “summary of summaries“. The negative effect of this is that sub-summaries cannot relatively weigh their importance against other sub-summaries - they are weighted equal whether they are or not.
Depiction of batching LLM medical record summarization
Hallucinations: Currently, it is not uncommon for hallucinations to still occur, especially in situations where output is asked for that cannot be generated easily from an input. There is not solid way to fully mitigate hallucinations from an LLM at this time.
Cost: Even with recent advancements in LLM tech (e.g. DeepSeek distillation), the cost of the use of LLM tech remains high. Even with batching, for a single task such as summarization costs are as high as $1 per 1000 pages for just summarization. 2024 tech was on the scale of $10 per 1000 pages. It is expected however for these numbers to further drop.
See our larger article on upcoming LLM tech for more information.
Conclusion
Medical record summarization is a transformative application of NLP and LLM technology in healthcare, offering the ability to condense complex, voluminous data into actionable insights. Tenasol’s multi-faceted approach—spanning extraction-based, abstraction-based, and OCR-enhanced methods—demonstrates the potential to optimize tasks like HEDIS quality measures, risk adjustment, and prior authorization.
While these advancements streamline workflows, challenges like input/output limits, batching, and hallucinations highlight the need for ongoing innovation. Emerging technologies, such as Donut models, offer promising cost efficiencies but come with trade-offs in accuracy. By tailoring strategies to client needs and embracing a balanced view of capabilities and limitations, Tenasol stands at the forefront of medical record summarization. As this field evolves, addressing hurdles with enhanced models and innovative batching techniques will be critical to unlocking the full potential of NLP in healthcare, driving better patient outcomes and operational efficiency across the industry.
Reach out to our team if you have interest in using these tools.