The note section method is
a text processing technique that uses the clinical note section
to provide context about the meaning of keywords
found in those sections. For example, let's consider medications, if you searched for
the drug name Codeine in a clinical note and found it in
the current medication section, you would conclude that the patient
has been prescribed the drug. However, if you found the exact same
name in the drug allergy section, you would probably
conclude that the patient is not being prescribed the drug. Similarly, the diagnosis hypertension
appearing in past medical history, has a very different
meaning than if it were found in the family
medical history section. When hypertension appears
in past medical history, you know the patient has hypertension. But when it appears in family history, you only know that someone in
the patient's family has hypertension, but not necessarily
the patient themselves. In each of these cases, the actual text found is identical, but the context and meaning of
the words are completely different. Let's talk a bit more about how to implement this technique with an example. Our goal is to identify patients who experienced muscle pain
while taking a statin. Statins are a class of
medication that are taken preventively to lower cholesterol and reduce the risk of
cardiovascular disease. Muscle pain or myopathy, is one of the most common
side effects and is one of the major reasons why patients
stop taking their Statin. The first step to solving
this problem is to identify which note types and what sections in those not types would have
the information we need. Oftentimes, drug side effects
that cause the patient to stop taking a drug are recorded in
the medication allergy section. Even if the reaction
isn't a true allergy. We know that history and
physical notes as well as things like admission
and discharge summaries, all tend to have
medication allergy sections. Now that we've identified
our note and section types, we can search the notes for section
headers that mean medication allergies. Our goal is to find the section and
extract all the text in that section. There are two general
approaches to this task. The first method, looks to see if
there's any standard section delimiter. For example, if all sections
start with a new line. If the notes are well formatted
and have this kind of structure, we can use that delimiter to split
the node into the individual parts. The second approach is more common, because many notes lack the consistent structures
needed for the first method. In this approach,
the first step is to identify all the different terms used in the dataset to indicate
the same note section. For example, sections labeled
drug allergies or even just allergies, will all mean the same thing. Once that list has been created, you'll want to look at the section header immediately following
the allergy section. Again, because each note may
have a slightly different order, you'll have a few different header types
for that next section. But once you have a list of
all those next section headers, you can easily extract
all of the information between the allergy header
in the next section. Regardless of which
extraction method you use, once you have that allergy
section extracted, you search for your keyword of interest. In this case, the stat name, including both the generic
and the brand name. For example, we would
search for atorvastatin and lipitor, simvastatin and zocor. We can also use the class name stan. If any of these drug names are present, then we can initially call them a case. They have an allergy to a Statin. However, many allergy sections actually list the specific reaction
caused by the allergy. In the case of statins, in some patients the Statin
actually affects liver function. This calls for an extraction task, pulling out the reason
listed for the allergy. In a lot of cases the reason is
stored between two parenthesis. Once we have extracted the text, we can search if the reason list
has something to do with muscle pain or
the clinical term myopathy. Sometimes you may find that
it's actually easier to just exclude all phrases that
don't indicate myopathy, than to pull out all the different ways the muscle pain reaction is recorded. If you have a record where
no reason has given, you can either ignore it and say
you don't have enough information, or conclude that because muscle pain
is the most common reaction, then it probably means they
had a muscle pain reaction. As you can tell, this approach does
require a bit of experience in understanding clinical note types and the types of sections
found in each note, as well as the meaning
of those notes section. Well, we'll cover this information
for a number of common note types. This is a case where having
a clinician as part of your clinical data science team
can make a huge difference.