By Eric Fischer, Senior Product Manager, Data Solutions
In part 2 of our series on how Accruent Data Insights works and how it helps the healthcare technology management (HTM) profession, we focus on the standardization of medical device categories.
The process of organizing medical devices into categories seems trivial – until you realize how fragmented and massive the medical device industry is. Mature categorization schemes such as UMDNS, GMDN and the FDA Product Code Classification are sophisticated, but without boundaries and strategies to map to these schemes you risk having as many categories in your database as there are device models themselves. Furthermore, consistency in applying the classification schemes is often lacking.
As a result, categorization of medical equipment has evolved into a problem area for many healthcare technology management (HTM) leaders. This leads to substantive, negative impacts on maintenance processes, planning processes and overall technology management in hospital systems.
Accruent Data Insights addresses this issue by standardizing medical device categories so that your organization’s data is more consistent and useful, no matter which categorization scheme you follow.
Using Artificial Intelligence to Standardize Categories
Accruent looked to modern big data and artificial intelligence strategies for a solution. In my previous post, I covered how Accruent Data Insights standardized all the medical device models in our data warehouse. Once that process was done, Accruent could view the categorization data for any given device model in unique ways.
One way of thinking about categories is to consider how a computer would look at them, which is by treating the categories as numbers. Yes – strings of words can actually be converted into numbers. This technique is a part of artificial intelligence called word embeddings. Word embeddings, or word vectors as they are often called, have been critical to making recent advances in AI, particularly in the field of natural language processing (NLP).
The most basic type of word embedding is called a word count vector. In extremely simple terms, it works by counting the number of words in a vocabulary, looking at each word’s frequency, and then assigning each word a vector. The higher the vector, the more frequently the word is used.
Below is an example of one for the “BD Carefusion Alaris 8100” infusion pump, based on the massive data about this device in the Accruent data warehouse:
[('PUMP', 0.2662337662337662),
('INFUS', 0.2077922077922078),
('MODUL', 0.11688311688311688),
('PATIENT', 0.032467532467532464),
('LARG', 0.032467532467532464),
('VOLUM', 0.032467532467532464),
('PURPOS', 0.025974025974025976), …
The word count vector above tells us that most of the time the item is categorized using the word “pump.” “Infusion” is a close second. (I will explain shortly why the root word “INFUS” is used instead of “INFUSION.”)
From a human standpoint, we know this is an infusion pump. The cool thing is that a computer program can reference the word vector to make educated predictions of what a good category match would be. When the computer program compares the word vector to the FDA standard categories list, the computer selects the FDA category of “PUMP, INFUSION” (not surprisingly!) as the best match. The word vector also tells us it is modular and a large volume pump, which are both shockingly accurate.
Eventually, if you were to look at the entire word vector, the occurrences of relevant words would drop off significantly and have little value. By design, they have low numeric values tied to them, and as such the word vector also sees them as irrelevant.
Back to how I mentioned “INFUSION” was shortened to “INFUS” in the word vector: Accruent decided to use two tactics common in this space called lemmatization and stemming. This helped with word vector creation because we would often see “pump” instead of “pumps” or “radiation” instead of “radiology” in the customer’s CMMS system.
Good Categories Lead to More Good Data
While the technology applied to the huge dataset worked well, the result was not perfect. So next, Accruent leveraged the FDA categorization as a check against what the computer declared was appropriate. We also performed significant manual human review of the results to improve the accuracy.
Once we had confidence in our categories and the ability to map to category schemes, we decided to automatically map all devices to the AHA. This helped us get a benchmark of useful life for each medical device and was used as part of the life expectancy intelligence, which will be discussed in a future blog post article in this series.
In addition to better life expectancy data, we were able to use the category data to find outliers in prices and reliability metrics based on the category level insights. So having good categories made the Accruent useful life, reliability and pricing data better as well.
How Accruent Data Insights Can Help You
We often hear customers struggle in their efforts to map to sophisticated categorization schemes. The good news is that Accruent Data Insights does not offer any new standard for you to follow; it simply helps you categorize using whichever standard you choose – which often is your own standard. And another plus is its Data Standardization report shows you how your data can be more consistent in its own categorization methods.
Accruent Data Insights is automated, so getting useful results is quick and painless. Even simply mapping to AHA to get life data you can act on is a breeze. If you’d like to learn more, please reach out. Ask your Account Executive for a free data cleaning report card so we can show your how this technology can help you. As always, we can be reached at sales@accruent.com or by visiting Accruent Data Insights to request a demo.
Looking for the full Data Insights Deep Dive weekly series? See:
.svg){kind=small}
