✨ AI Extraction Guide

Flinn's AI Data Extraction Feature

Turn 1 hour of manual extraction into 10 minutes of AI input review

Before Flinn AI, the tedious literature review process led clinical writers to include as few papers as possible and extract only the minimum needed to draft.

Today, those time consuming steps can be delegated to Flinn AI. As a result, the range of publications assessed grows and the completeness of data extraction increases. This requires a mindset shift: clinical writers can now expect large volumes of high quality, structured data as systematic review outputs, which can later be filtered or transformed as needed. This is how it works:

Let AI autofill your extraction template directly
1-click trigger for all extraction fields
Review the proposed values and edit them if needed
Export with confidence - all extracted values are automatically documented in your Excel report

🔧Use Pre-Defined Fields

We’ve partnered with clinical experts to develop 13 ready-to-use AI prompts, now available from version 2.9 onwards.

These pre-defined extractions, such as Population Sex, Complications, Study Design, and many more, can now be applied instantly to any new or existing text fields, saving time and boosting consistency across your reviews.

➡️ To get started, head to Literature Settings > Extraction and explore the new options today.

✍️ Create your own AI fields

In addition to using predefined fields for AI data extraction, you can create custom fields by providing your own prompt, allowing you to tailor the extraction to your specific needs.

AI extraction feature disclaimer: Flinn’s current AI extraction is built to support up to 90% of extraction fields. This version prioritizes simplicity to make extraction accessible to all manufacturers. The remaining 10% of advanced fields will be addressed with an individualized approach in a future release.

Go into Settings > Literature > Extraction
Click on Default Template
Create a Group and add a field
Select Text and Custom prompt
Name your field and describe what you want AI to extract

💬 Frequently Asked Questions

What is the impact of the title vs the description on the extraction quality?

A title is required to create an extraction; a description is optional. A title alone can work for simple, obvious fields such as publication authors or publication year, where the AI already understands the concept.

For more complex extractions, the description determines quality. Including clear examples in the description yields the highest accuracy.

What is a good example to add to my description?

A good example pairs a sample input with the expected output. Below, show the original sentence from the publication and the exact element that should be extracted.

**Example 1 Text**: Histological evaluations were performed using a light microscope (Eclipse ME600; Nikon, Japan) and histomorphometrical data was analyzed by an image analysis software (Image J ver.1.43u; National Institutes of Health).

**Example 1 Expected Result**: Nikon Eclipse ME600

Can the prompt be prompted with 'Do not' or 'Don't'? How to achieve this?

You can include exceptions in the description. We strongly recommend adding examples that illustrate the exception, so the AI can infer and generalize the rule.

Extract all names of medical devices that are used in the current study.
Only physical devices are relevant that are used in any kind of study context in the paper at hand (e.g. scalpels, X-ray apparatus, disinfectants, dressing material). List not only the device name but also the manufacturer name, e.g. Philips Glucosemaster 3000. Do not extract any software that was used to generate the results or to analyze the data (e.e. IBM SPSS)

**Example 1 Text**: Histological evaluations were performed using a light microscope (Eclipse ME600; Nikon, Japan) and histomorphometrical data was analyzed by an image analysis software (Image J ver.1.43u; National Institutes of Health).

**Example 1 Expected Result**: Nikon Eclipse ME600

What are the sections considered by the AI?

Currently, the AI reviews the entire paper. It searches first in the main body (methods, results, conclusion), then in the introduction and abstract, and finally, if nothing found, in the references.

Can I use the extraction to compare the information in the text to defined criteria/thresholds for me (and therefore support me in the evaluation)?

It can work, but it requires examples. Be explicit about what “compare” means: include an example with the extracted value, the ground-truth value, and the rule for how they should be compared.

Can the extraction perform (simple) calculations or will it "just" extract what is there?

It can compute a minimum or maximum when there aren’t too many data points, and it can do simple addition or multiplication with two numbers. For these cases, name the field “Minimum” or “Maximum” and use a specific field rather than a generic calculator.

Complex calculations on large tables (for example, a mean) are not supported yet.

Does the extraction "know" what's the Device of Interest in my search and make the extraction based on this? Or is it general?

The extraction works in a general way - it won’t target a specific device unless you ask it to. The best approach is to capture all details and then filter for your Device of Interest later when you use the Flinn Library or start writing your report.

If needed, you can narrow the description to a single device, but that makes the template device-specific and usually means maintaining multiple templates. For most teams, staying objective (not device-specific) in the template helps maintain flexibility and scalability.

How to use extraction template the best way?

Templates are intended for extracting data that will later be used in very different contexts, such as reports types (SOTA vs. device-specific) or by teams working independently. If you create templates whose extracted data will later be used together to write or report, you will need to merge the data, which is not recommended. Always keep in mind that extraction aims to be as exhaustive as possible, and filtering happens at a later stage. For example, if you need a patient number for specific devices, first extract all devices mentioned with all population numbers associated. When you search the data or use it to write, you can then apply a filter to consider only the relevant part of the extraction.

Is this feature validated?

Yes, the extraction feature has been validated to ensure accuracy and reliability. It has undergone structured testing with representative data and review against predefined benchmarks, so you can trust the results as a solid basis for your work.

👤Need help?

Book a meeting with your CS expert or write us an email on support@flinncomply.com