While most people have heard of generative artificial intelligence (AI) following the global rise of Chat-GPT, fewer people are firm on the details. What is extractive AI, and how is it different from abstractive AI? Where does generative AI come in? Most importantly, how can these tools be used for medical records review?

For the purposes of this example, let’s consider a patient’s medical file as a library: extractive AI tools act as a librarian, a custodian who helps the user collect what they need. Abstractive AI tools use the library too, but these act more like a news reporter, pulling the facts together into a narrative form. Both abstractive and extractive AI can be valuable in medical records review – so long as it is an appropriate situation for each to be used.

Extractive AI: the “librarian” of your records and files

Like a librarian, extractive AI carefully searches through a document (its “library”). Without changing key phrases or sentences from the original wording, it lifts the most relevant pieces of information, and presents it to the reader to use. Extractive AI tools work with objective data – including dates, titles, hospitals, patient numbers, or other details – to inform the machine learning tool of what the document could be, and where information is stored within it. Extractive AI then finds these important pieces in the original text, selects them, and brings them together in a format the reader can use quickly.

A librarian might bring some relevant books to you based on the topic, but they don’t write the report – you do. Just as the name implies, extractive tools do not create anything unique. Extractive AI simply pulls information out of the document for the user. This function can be extremely useful in medical record summaries where it’s essential to protect the accuracy and fidelity of the source. Wisedocs’ medical summaries are an example of extractive summarization: pulling the important sentences out of a report, for example, without generating any content by itself.

Abstractive AI: reporting the “headlines” of the medical record

In contrast, abstractive AI approaches content more like a reporter. It not only extracts key data, the way extractive AI tools do, it is also able to interpret, rephrase and generate text that summarises and humanises what it “reads.”

This method can save the human user a lot of time. Imagine a news story you read in the paper about a car accident. Since you don’t need to report on the details, you trust what the reporter gives you: the who, what, when, where, and how. If the claims file for the accident makes its way to your desk, you’ll have your own work to do. For now, the news story has given you a pretty good overview.

Abstractive AI tools provide a similar function for humans to use. AI is able to work through massive volumes of information at machine learning speed. It can simplify documents and ideas without losing the meaning that they hold, and summarise these documents for the user at a glance. Wisedocs interactive timeline view is a key example of an abstract AI tool, allowing users to see an overview of the claims case and gain a comprehensive understanding of the patient’s health records. 

In this way, abstractive AI “reports” on what it reads, and stays true to the source. However, it is important to note that this kind of AI requires proper oversight to make sure nothing is lost and that the most accurate information is presented to the user.

‍

When to use extractive vs. abstractive AI?

Imagine you asked two first year law associates to summarise a report. One of your associates, E, highlights the most important sentences and hands them over – now you have a highly accurate collection of key points for your review. Your other associate, A, gives you a quick rundown of the report in the elevator – in just a few minutes, you’re ready for the client meeting you’re heading into. 

The hypothetical lawyer in this example would need to rely on E’s highlights when preparing the case for the trial – but for the purposes of getting prepped for the meeting, A’s work saves plenty of time. 

Both extractive and abstractive AI tools are helpful in reducing manual workload and speeding up workflows. However, there are times when one is more appropriate than another.

• Extractive AI tools are most reliable when it’s important to preserve the fidelity and accuracy of the source. When reducing the preparation or paperwork time of a human medical expert, for example, an extractive AI tool can prepare and organise a file in a fraction of a human’s time – while still preserving the details. 

• Abstractive AI gives the viewer a ‘birds eye view’ of the file – a quick, easily digestible narrative that tells the human user exactly what they’ll expect to see. This is useful when a party to the claim wants to see an ‘at a glance’ snapshot of a patient’s health history and what the claim entails. 

Extractive AI tools are the most reliable and accurate reflection of the original source data. Abstractive AI tools give the user fast, easy to digest information that they can use. Both are sometimes necessary to handle a claim, but it is extractive AI platforms that are necessary to protect the fidelity and privacy of the data you use.