Medical Publications Search

---

A Fortune 500 medical devices company faced challenges related to product approvals and recertifications. The issues were rooted in regulations requiring systematic medical literature review and analysis to find clinical evidence regarding the use of their devices in appropriate clinical scenarios. Such clinical evidence reporting is needed for not only initial product approval, but also for ongoing product recertification (potentially requiring that multiple reports be submitted per year, per device). The company employed large teams of Medical Writers who spent countless hours performing keyword searches for relevant medical publications, then filtering and assessing within search results the appropriateness of found clinical evidence for Health Authority reporting. 

The process for searching publications – and then annotating relevant results for subsequent use for detailed analysis for evidence usage – was analyzed with clinical leaders and Medical Writers. The main request centered around the notion that “Once I have finally found the type of publication I was looking for, I want to find other publications ‘like it’”. Keyword search is unable to achieve this, and currently available AI search techniques also have limitations (including AI model bias) in their ability to find “articles like this one”. So with this usage paradigm in mind, the data science team endeavored to innovate a new ML search method.

A new technique for ML search was invented that allows users to take one or more exemplary publications and instantly train a domain-agnostic, search-session-specific ML search engine to go find other publications like the exemplar(s).  Keywords were used to define what we call the "Search Universe". The Human-in-the-loop feedback (implicit + explicit), within the keyword and Boolean-defined "Search Universe", enabled this breakthrough AI-on-the-fly content discovery technique.

Search result annotators helped organize results for clinical reporting purposes.

The ability to search for content using not just keywords – but using an exemplary piece of content – could enable hard ROI while also making possible other unexpected benefits:

• Cost-Savings: By allowing investigators to more quickly find what they are looking for, the medical devices company estimated it could save over $10M USD per year in Medical Writer consultant fees. 

• Performance & Consistency: Because keywords and complex Boolean strings are not required for “pointing the ImplicitSearch™ engines in the right direction”, inexperienced investigators can be more effective, with search efforts across teams being more consistent.

• Expert-Initiated Investigations: Experts can initiate investigative sessions by training an ImplicitSearch™ engine on an exemplar, then handing off that search session to others who are less experienced in the domain. This “flips the script” on this type of work (it’s commonplace for staff or clerks to start an investigation then send their findings to a senior-level expert who determines relevancy to the task).

• The solution could serve as a literature “Discovery” tool because the content used to train the ImplicitSearch™ engines needn’t be existing publications; investigators could instead use elaborately crafted hypotheses to train the engines – the results from which being seen as “evidence of those hypotheses”.