Unpacking Averages: Analyzing Descriptions from Cardiovascular Device Adverse Events

Whereas this column usually makes use of knowledge visualizations you’ve most likely seen earlier than, I wish to introduce one which maybe you haven’t.  That is within the realm of textual content evaluation.  When FDA knowledge, there are quite a few locations the place essentially the most fascinating info isn’t in an information discipline that may be simply quantified, however slightly in narrative textual content.  Take, for instance, Medical Device Studies of adversarial occasions, or “MDRs.”  Whereas we will do statistical evaluation of MDRs exhibiting, for instance, which product classes have essentially the most, the actually fascinating info is within the descriptions of the occasions.

Why can we care?  These targeted on product high quality wish to be taught from historical past, and the way higher than from everybody’s historical past, not simply your individual.  We will use the MDR knowledge to search out out product experiences with broad or slim product classes.  Within the 12 months 2020, for instance, there have been over 1.5 million such stories.

On this month’s column, I’m going to concentrate on cardiovascular gadgets.  I may simply as simply concentrate on software program, or all implants or orthopedic gadgets or every other time period that may cowl a broad vary of merchandise.  As a result of these things isn’t frequent, I’m going to incorporate the methodology first, after which the visualization.


I’m utilizing what is thought in knowledge science as subject modeling to extract info from the occasion descriptions in MDRs.  Matter modeling is an strategy that enables us to see the large image from an extended record of paperwork.  Primarily, subject modeling seeks to establish frequent matters mentioned in a corpus of information.

Whereas the output form of appears like English, it additionally form of doesn’t.  The algorithm is in search of phrases which might be used collectively regularly.  The algorithm strings these phrases collectively based mostly on statistical significance, not based mostly on how an English main would write them.  Additional, to assist the algorithm acknowledge that the phrase “implant,” “implanted,” “implants,” and “implantation” are all fairly comparable, we cut back every phrase to its stem.  Within the chart, you will note all variations of the phrase valve as “valv”.  It takes some getting used to.

As well as, as a result of context is often vital to grasp how phrases are used, we search for phrases, on this case two phrases which might be usually strung collectively.  We may use nevertheless as many phrases in a phrase as we wish, however extra phrases means much more pc sources.  On this case, two-word phrases appears to work adequately.

This train is a mix of artwork and science, and one of many judgments to make is the variety of matters to think about.  We usually make that call based mostly on what we check with as coherence, which is a statistical measure of what’s a logical precept: how properly the phrases go collectively.  We wish to discover matters which might be significant to people.  One other space for utilizing judgment is eliminating phrases we don’t care about as a result of they’re too frequent and uninformative.  I eliminated phrases reminiscent of “complaint” and “patient,” as a result of they have been in lots of the occasion descriptions and didn’t add any significantly helpful info.

From a technical standpoint, for these , I’m utilizing a particular method referred to as Latent Dirichlet Allocation, a type of unsupervised studying, carried out by means of the Python library genism.

For cardiovascular gadgets, I assumed it is perhaps fascinating to check the matters that MDRs included in 2010 with these from 2020.  I needed to see how a lot change there is perhaps over time.  Within the 12 months 2010, there have been over 45,000 MDRs for cardiovascular gadgets, and in 2020, there have been nearly 85,000.


A phrase about studying these charts.  These are what are known as warmth maps.  The colours correspond with the depth or worth of a selected phrase, on this case to the subject.  The darker the colour, the extra vital the phrase in characterizing the which means of the subject.


An knowledgeable in cardiovascular gadgets will undoubtedly be capable of provide a lot larger insights into what these knowledge imply, nevertheless it’s really fascinating that over the course of 10 years, the software program discovered what turned out to be many very comparable matters.  The order modified.  However lots of the matters appear to be fairly comparable.  For these within the business, which may be a bit miserable because it suggests a scarcity of progress in fixing frequent issues.

Balloons rupturing nonetheless appears to be a difficulty.  Guidewire ideas nonetheless appear to be a difficulty.  Battery issues nonetheless appear to be a difficulty.  However we even have some new points, reminiscent of issues with a floor cooling gadget (Arctic Solar®) for therapeutic hypothermia following cardiac arrest.

This method of subject modeling will be utilized as broadly or as narrowly as we want.  It may be very useful when doing development evaluation the place the precious underlying knowledge is in massive volumes of textual content slightly than structured knowledge.  In future columns, I’ll dig into different sources of regulatory textual content to see what info we will mine.

Source link