AI data extraction for literature reviews: what's new in NOÉ this October
6 Min Read
Product Updates
In this release: a redesigned data extraction, reports covering up to 200 articles, and automatic citation checks.
By Bopha Suybeng, PharmD, founder of NOÉ. Twelve years of experience in clinical evaluation of medical devices.
October was a busy month at NOÉ. We released three versions: 4.10, 4.11 and 4.12. They share one goal. Save you time on your clinical evaluation literature reviews, with no compromise on reliability.
In this release: a redesigned data extraction, reports covering up to 200 articles, and automatic citation checks. Here is what changes for you.
Why these changes
A literature review for a CER or a state of the art takes weeks (see our comparison of manual and NOÉ review timelines). Most of that time goes into data extraction. You read each article, find the figures, and check where they come from.
Your feedback showed us two limits. First, some article types misled the AI. Second, our reports were capped at 50 articles. Yet many reviews include 100, 150, sometimes 200 articles.
Your feedback shaped the October releases directly. Thank you.
Data extraction: NOÉ now recognises each article type
This is the biggest change this month. NOÉ now identifies the type of each article before extracting its data.
The problem: other people's figures
A narrative review cites many studies. Its figures are often not its own. They come from the articles it cites.
An AI can get this wrong. It may credit a result to the article it is reading, when the result comes from another study. Or it may link the result to the wrong reference. For a clinical evaluator, this is a serious error.
The solution: extraction tailored to each article type
NOÉ now tells apart several families of articles:
systematic reviews and meta-analyses;
primary studies;
narrative reviews;
classifications, supplements and other formats.
Each family gets its own extraction instructions. Figures are linked to the right source. A cited result stays credited to its original study.
Measured results
We tested this change on our internal control set. It contains deliberately difficult articles: narrative reviews, classifications, supplements. Overall extraction quality rose from 73% to 90%.
This improvement is already live for all NOÉ users.
Other extraction improvements
Progressive display. Results appear as they are ready. You can start working without waiting for the end.
Clearer tables. Extracted data is presented more cleanly.
More flexible clinical risks. They no longer need a citation to appear.
Better handling of short PICOS. Criteria of one to three words now work correctly (a reminder on the PICOS method in clinical evaluation).
Custom questions. Long titles no longer break the layout.
Literature review reports covering up to 200 articles
Until now, a NOÉ report covered 50 articles at most. You can now generate a report on a review of 150 to 200 articles.
How it works
NOÉ no longer processes the review in one block. The report is built in several steps, each one checked:
NOÉ splits your outline into sections.
Each section receives the data categories it needs.
Articles are processed in batches, then summarised.
Each section is written from these summaries.
The result: no information is dropped to fit a limit. Each section only receives the data it needs. The report becomes more precise.
Generation takes time, by design
A large report takes time. Allow 30 to 40 minutes for about sixty articles.
NOÉ warns you before it starts. Sections appear as soon as they are ready. An introduction arrives within minutes. A performance section, full of tables, takes longer, just as it would for a human writer.
Expert Check: automatic citation verification
The longer a report, the higher the risk of citation errors. This is true for an AI. It is true for a human too. By the 56th article, two references are easy to mix up.
Expert Check automatically verifies figures and citations. It focuses on the most sensitive parts: tables, discussion and conclusion.
Here is an example. The report says "3%" and cites article 8. Expert Check looks for this figure in article 8. If it cannot find it, it flags it in orange. The figure may still be right, for example if it was calculated. But you are prompted to check.
We made a deliberate choice. Expert Check does not rewrite the text. Automatic rewriting could introduce new errors. It works like a reviewing colleague: it comments, you decide. On the role of the expert alongside AI, see also our article on double human screening in an AI-assisted systematic review.
When you export to Word, these flags appear as comments. You handle them like any other review.
This has been our belief from the start. AI speeds up the work. The expert stays in control. This is what Regulation (EU) 2017/745 (MDR), the MEDDEV 2.7/1 rev. 4 guidance and your Notified Body expect: traceable, verified data.
Coming next: NOÉ Elephas
October laid the foundations. The next major version, NOÉ 5, arrives in the coming weeks. We have named it Elephas: powerful, intelligent, and with an excellent memory.
What's coming:
Review updates. Create a new version of your review without starting from scratch. NOÉ keeps track of every version. You always know where each article comes from. The PRISMA diagram combines all versions.
Statistics and data visualisation. Charts and summary tables: study type, population, level of evidence, device, outcomes. Add or remove columns as you need.
Custom data extraction. Extraction built around your own templates.
A new materiovigilance module will follow.
Frequently asked questions
How many articles can a NOÉ report cover?
A NOÉ report can cover up to 200 articles. Articles are processed in batches and summarised. No data is dropped to fit a limit.
How does NOÉ avoid errors in narrative reviews?
NOÉ first identifies the article type. A narrative review is handled differently from a primary study. Cited figures stay credited to their original study.
Is the extracted data traceable?
Yes. Each data point is linked to its source article. Expert Check flags the figures to verify. You keep the final decision, as your Notified Body expects.
Is NOÉ suitable for CERs and state-of-the-art reviews?
Yes. NOÉ is built for literature reviews in clinical evaluation reports (CER) and state-of-the-art (SOTA) documents. It follows the PICOS and PRISMA methodology.
Try the new features
All October improvements are already available in your NOÉ workspace. For an overview, discover all NOÉ features for literature reviews. Not using NOÉ yet? Book a NOÉ demo. We will show you what NOÉ can do on your own device.
About the author — Bopha, founder of NOÉ. Bopha created NOÉ to support regulatory affairs and clinical evaluation teams in their systematic literature reviews (CERs, state-of-the-art reviews, MDR/IVDR). Her belief: AI speeds up the work, the expert stays in control.