Search

Effective Medical Literature Monitoring Workflows with AI

The monitoring of scientific literature for adverse events is a high volume activity requiring more time from specialists every day. At biologit, we’ve set out to tackle this issue by developing an end-to-end screening solution with AI at its core, always working closely with domain experts in pharmacovigilance to make sure we’re addressing the real needs of users.


We have previously discussed how AI models can act as an effective pre-screening mechanism to categorize incoming abstracts according to their relevance for safety information. In this post we will discuss the features of biologit MLM-AI that apply artificial intelligence to help professionals focus on the articles that matter.


Abstract Tags

In MLM-AI, every abstract resulting from a search is tagged with the prediction results from a number of NLP and machine learning models. Each tag was created with the literature screening workflow in mind:



Of special importance is the Suspected Adverse Event model, which is trained to identify if the abstract describes a safety event (in humans), or if one may be found in the full text of the article. The main idea of this predictor is to facilitate filtering of articles without any safety data while ensuring important safety information is not missed - hence the prediction of a “suspected” event. The same idea of a "suspected" article is seen on other suspected AI models available in MLM-AI for animal/in-vitro studies and cases.


This means a “suspected” prediction is broader than say, a confirmed event, as illustrated in the chart below: out of all retrieved articles in a search, a subset are categorized as suspected AE, and this subset is where ICSRs (or other relevant safety-related articles) will most likely be found.


Lets see now how MLM-AI can take advantage of these AI tags:


AI Tags for Ranking, Fast Screening and Filtering

Once search results are tagged by the models, how can we use them? There are three different options that accommodate the specific levels of process oversight users may need.


Search & Rank

Tags are a useful mechanism for prioritizing the workload according to the most likely articles. For instance, in an ICSR workflow it would be useful to prioritize articles where identifiable patients are found first, followed by articles flagged as Suspected AE.


From the Review summary page, the search box allows users to craft queries taking the tags into account. Results can also be sorted according to tags. The example below filtered results according to the “Suspect AE” tag.



Once articles are reviewed, it is also possible to focus on the quality control effort where the greatest payoff is. Users can perform QC on a sample of results according to a search. For example, QC could be done on only 50% of articles not classified as Suspected AE.


Noting that search and sampling selections are remembered on the detail screening and batch screening pages, allowing users to do screening only on the chosen articles:



Batch Screening

Another option is to perform a faster type of screening on articles not likely to be relevant. For instance, if the screening workflow typically excludes animal and in-vitro studies, screening these articles can be done by (1) searching by tag and (2) performing faster “batch” screening.


In batch screening abstract and title are presented in sequence according to a specified criteria and allows users to save the screening decision once across all selected articles:


Ultimate Automation: Tag Filtering

Finally, incoming results can be automatically filtered according to tags. Auto-exclude settings are applied when configuring a new search:



Auto-filtered results already appear as reviewed, and can still be audited/quality checked just like any other screening decision.


Safe and Auditable Processes by Design

Whatever the level of automation customers chose, MLM-AI will always present the entirety of search results and any automated decisions in an auditable activity log. Nothing is ever hidden from users. All automated decisions can undergo normal quality control processes, just as would be the case when quality checking human decisions.


Conclusions


In this post we discussed various productivity features available in biologit MLM-AI that enable more effective screening workflows in pharmacovigilance. Users can choose the right level of automation and process oversight that meets their requirements, all within a fully auditable environment.


AI productivity gains can save professionals time and effort. They can also lead to results of higher quality: by focusing time on results that matter, more articles can be screening leading to strategies that more comprehensively search the literature. In this blog post we used MLM-AI to assess the benefits of searching open access literature for finding new events with positive results.


Learn More


To learn more about biologit MLM-AI, check out the Concepts Guide in the product documentation or get in touch with us at sales@biologit.com.