ezDI Story - How Healthcare Data Turns Intelligent?
©
ezDI, Inc
2017 — 2018
HL7
CONNECT
Unstructured & Structured Data
HIPAA
HIPAA Compliant Cloud
TM
Data Normalization
CLINICAL
Clinical data normalization or Pre-processing identifies and streamlines all inconsistent
content to provide a uniform input to the Natural Language Processing(NLP) engine.
Know more
Knowledge Graph
CLINICAL
It is a large collection of all terms available in the medical domain.
All these concepts are arranged in a hierarchy of concepts.
Know more
Natural Language Processing
Natural language processing (NLP) is a field of computer science, artificial
intelligence, and linguistics concerned with the interactions between
computers and human (natural) languages. There are 14 main components in
NLP.
Explore NLP Components
ezHealth
TM
HEALTHCARE
HOW
DATA
TURNS
INTELLIGENT
ezDI Cloud Platform has been designed from ground up with HIPAA
compliance. We employ an information security team that monitors
our data centers and applications 24X7X365. Not only do we ensure
that all HIPAA guidelines are met, such as Technical, Physical,
Administrative safeguards, but we also follow all current
healthcare industry best practices for secure transmissions
and sessions, network protection, environmental controls,
and more, mitigating the risk of a data breach.
While improved productivity and cost savings have long been at
the top of the list, more and more hospitals are choosing ezDI Cloud
Platform for its security and reliability benefits.
ezDI's HIPAA Compliant Cloud
Natural Language Processing(NLP) is a complex combination of
multiple components. Accuracy of NLP depends on the efficiency of
each individual component, but more importantly it depends on the
quality of documentation. The better and uniform the input data,
the better the output accuracy.
The biggest challenge in processing healthcare data is the
inconsistency of documentation. Each physician has his/her own
style of grammar, punctuation, abbreviations/slangs etc.
Clinical data normalization or Pre-processing identifies and
streamlines all inconsistent content to provide a uniform input to
the Natural Language Processing(NLP) engine.
Clinical Data Normalization is the answer
Providing uniform input to NLP, but how?
14 Components inside ezDI's NLP
Identifying the correct boundary of a sentence.
Sentence boundary detector identifies that the boundary of
the sentence that it does not end at b.i.d. but at “frequently”.
e.g. The patient takes aspirin 325 mg b.i.d. frequently
since May.
Sentence Boundary Detection
14 Components inside ezDI's NLP
Unstructured clinical data primarily consists of healthcare
documents transcribed based on physician dictations
where the physician dictates the entire encounter he
had with a patient.
There could be different elements of this document which
are classified into various sections like History of Present
Illness, Past Medical History, Past Surgical History, Current
Medications, Allergies, Physical Examination, etc.
These sections may be documented in different forms in
different hospitals and different physicians.
e.g. History of Present Illness may be documented as HPI,
Present Illness, Brief History etc.
Section Detection
14 Components inside ezDI's NLP
Tokenizer
This component breaks down a sentence into its constituent
tokens so that the next component which is the Part of
Speech tagger can assign a POS tag to each constituent
token.
e.g. The tokens are generated as
Word Token 1: The
Word Token 2: patient
Word Token 3: takes
Word Token 4: aspirin
Number Token 1: 325
Word Token 5: mg
Word Token 6: b.i.d.
Word Token 7: frequently
Word Token 8: since
Word Token 8: May
Symbol Token 1.
14 Components inside ezDI's NLP
POS tagger tags the different grammatical components of a
sentence based on Part of Speech like Noun, Verb, Adjective
etc.
e.g. The patient takes aspirin 325 mg b.i.d since May.
The/DT
patient/NN (Noun)
takes/VBZ (Verb)
aspirin/NN (Noun)
325/CD (Cardinal Number)
mg/NN (Noun)
b.i.d/NN (Noun)
since/IN (Preposition)
May/NN (Noun)
./.
Part of Speech (POS) Tagger
14 Components inside ezDI's NLP
Chunker breaks a sentence into different phrases like Noun
Phrases (NP), Verb phrases (VP), prepositional phrases (PP)
Adjective Phrases (ADJP) etc. as per syntactic rules.
e.g. “Sensation is intact to light touch in both lower
extremities” is broken down.
The chunker output would be:
Sensation (NP) > is (VP) > intact (ADJP) > to (PP) > light touch (NP)
> in (PP) > both lower extremities (NP)
Chunker
14 Components inside ezDI's NLP
Parser establishes relationships between different phrases in
a sentence following phrase structure rules as defined in
syntactic English grammar.
e.g. “Sensation is intact to light touch in both lower
extremities”
The parser gives the output as show in the visual.
Parser
S
S
ADJP
JJ
PP
VP
.
.
NP
VBZ
NNP
Sensation
Is
Intact
TO
NP
14 Components inside ezDI's NLP
This component establishes relationship between different
words in a sentence.
e.g. “Sensation is intact to light touch in both lower
extremities”
As shown in the visual, the dependency parser relates
intact
<-->
sensation, touch
<-->
light and so on
Dependency Parsing
14 Components inside ezDI's NLP
The dictionary look up component of NLP maps the concepts
identified from the document against concepts present in the
ontology which is a comprehensive collection of medical
concepts classified into their types. Based on this look up, it
assigns tags of disease (problem), procedure, anatomical
structure etc. to the concepts.
e.g. “The patient takes metformin for his diabetes”
Metformin is tagged as a medicine and diabetes as a
disease or problem.
Dictionary Lookup Process
Medicine
Disease
14 Components inside ezDI's NLP
NLP has a built-in algorithm that establishes the primary level
of relationship between concepts such as anatomical structure
and problems or diseases, procedures and anatomical
structure and procedure and medical devices.
e.g. “The patient has intermittent pain located on the
left side of his chest.”
NLP can relate pain to the chest and identifies chest pain
although they are not co-located within the sentence.
Relationship Finder
14 Components inside ezDI's NLP
The UEI (Unique Entity Identifier) detection module uses
the relationships identified between different words in a
sentence to identify matching concepts in the ontology or
knowledge base to assign a unique identifier called a UEI
or Unique Entity Identifier.
e.g. “The patient complains of pain in the leg”
The ontology has a UEI for leg pain. The UEI detection
module identifies that “pain in the leg” is the same as
“leg pain” and assigns it the relevant UEI.
UEI Detection
14 Components inside ezDI's NLP
NLP has a negation-detection algorithm that identifies such
indicators to identify negation in sentences.
e.g. “There is absence of any cardiac enlargement”
Here the word “absence” indicates that cardiac enlargement
is not present or is negated.
Negation
14 Components inside ezDI's NLP
Temporal Status Detection is a component of NLP that
identifies the status of each concept with respect to its
temporality which is present, past, future etc.
e.g. “The patient has had an MI in the past”
The sentence is the past tense, detected by the
word ‘past’ and its mapping with ‘MI’.
Temporal Section Detection
PAST
PRESENT
FUTURE
14 Components inside ezDI's NLP
Modifiers are terms used to further describe the specificity of
concepts in a medical document. This component identifies
which concept with a modifier is related to and forms the
groups of words like ‘intermittent pain’ are marked as
modifiers for pain.
e.g. “The patient has intermittent pain located on the
left side of his chest”
‘Intermittent’ and ‘Pain’ words are grouped to specify
the concept.
Modifier Detection
14 Components inside ezDI's NLP
Drug Mention Annotator
There are various attributes associated with a drug (medicine).
The drug mention annotator identifies the various parameters
and establishes a relationship between the drug and its
parameters.
e.g. “The patient's Lasix was changed from 20 mg to 40 mg
tablets p.o. b.i.d.”
Drug: Lasix
Strength 20 mg, 40 mg
Route: p.o. (by mouth)
Frequency: b.i.d (twice a day)
Status: Change.
ezDI's clinical knowledge graph is a large collection of all terms
available in the medical domain. All these concepts are arranged
in a hierarchy of concepts. All concepts are primarily classified as
Diseases, Procedures, Medications, Medical Devices, Anatomical
Structures..etc.
Each of these concepts are further divided and subdivided into
further concepts. Moreover, each concept is related to each other
concept with a particular relationship.
Clinical Knowledge Graph
ezDI knows all the Medical Terms and their correlations
