Reviewing large amounts of medical texts requires a lot of repetitive work, searching for the relevant information in thousands of papers. Xtract AI built a solution to locate relevant information in unstructured data and extract it for review, cutting review completion time from years to hours.
Systematic literature reviews are a key component of evidence-based decision making. In practice, literature reviews are often expensive, slow, and time-consuming processes. In public health, this is in large part due to the massive amount of unstructured information that must be filtered through manually. Synthesizing key information from articles to create a systematic review can require a considerable amount of expert time and involvement. Completion times often exceed one year, and costs can run into hundreds of thousands of dollars.
Much of the work required when conducting a systematic review is repetitive and manual. As the literature in a particular field is rarely standardized in terms of format, an expert must read many articles with varying structures and layouts, when conducting the systematic review. It is common for the expert to apply inclusion criteria to determine whether to include an individual piece of literature in the systematic review. Ultimately, locating key information to determine whether a paper should be included in the systematic review is slow due to this lack of standardization. As the task of searching for key information must be repeated over a large number of papers, the work quickly becomes very repetitive and time consuming.
Systems that help guide one to relevant pieces of information hold the potential to reduce the amount of repetitive work involved in creating a systematic review. The solution we created predicts the location of relevant information in unstructured data. It does this by
- labelling relevant information in a variety of immunization-specific papers, such as study size and vaccine safety, carried out by domain experts
- training an AI system using data extraction and analysis to surface this information in previously unseen papers
- guiding the expert undertaking the systematic review to the location of the relevant information, instead of them having to search for the information manually
Computers have already made the process of searching for information in structured databases relatively painless for humans. Further to that, using AI for text extraction and analysis holds the potential to help do the same for unstructured data, by automating tedious, time-consuming work.
The AI solution decreased the time to undertake systematic reviews from years to hours, and as such saved other resources both in terms of money and people hours, freeing up the experts who were undertaking the reviews to perform more high-value functions. The amount of time saved by a researcher not needing to read every page of thousands of research papers is much greater than the time required to initially train the AI system.
By building out a user friendly interface, interacting with the AI solution can be made almost as easy as interacting with an Excel spreadsheet. The simple web interface shown in the prototype below could be used to query the system, producing the example search result shown below.