Medical Knowledge graph
ES, version 2025-10-03
- Problem:
- A large amount of biomedical knowledge is already
available but ln too many different formats and sources.
However this expertise is not always applied in practice.
- Objectives:
- A medical knowledge base suitable for decision support and
including quantitative evaluations.
- Approaches:
- Steering for decision support:
- Decision support require a very comprehensive
representation of the knowledge. Therefore graphs can
represent concepts and the very many relationships
between these concepts.
- Steering information intended to be process-able by
the decision support engine.
- Multi-directional navigation:
- From symptoms to potential "Problem".
- A symptom may leads to several hypothesis, with
their relative probability.
- Remark: What most matter here are "Issues" requiring
attention, but the name "Problem" is the most
intuitive.
- From a "Problem" to possible "Recommendations".
- Recommendation to seek more information, ask more
question, more physical examinations, more labs, ...
in order to increase or decrease the probability of
hypothesis.
- Recommendations to treatments.
- .....
- Concepts:
- Identification:
- In medical graphs, what most matter is a
"concept" rather than a "language". As far as
possible concepts and relationships should be seen
as logical notions, neutral from any spoken
language.
- Languages:
- Of course linguistic is necessary in order to
manage concepts. But it remains an approximation
due to cultural contexts and ambiguities.
- Medical concepts are in principle independent
from any spoken language. For example:
- The same symptom of "increased central body
temperature" as 发 烧, высокая
температура, " حمى",
fièvre, koorts, Fieber, fiebre,
... although not always exactly
one to one translations.
- The problem of gastric ulcer, as 胃
溃 疡, язва
желудка, úlcera
gástrica, Magengeschwür,
maagzweer,
قرحة
المعدة, ulcère
gastrique, ...
- As far as possible keep compatibility with
existing ontologies. At international level one of
the most comprehensive identification system seems
to be the ULMS, Unified Medical Language System, www.nlm.nih.gov/research/umls/
, including other ontologies and translated in
many languages. Anyhow in order to remain neutral,
a prefix indicating which identification system.
- However keep in mind that every existing
ontologies have been intended for a specific
goals, mainly for epidemiology, e.g. IDC from the
WHO, or insurance purposes, or for retrieval in
scientific literature. Decision support could
require a little different approach.
- Standardization:
- Raw data should be transformed in standardized
format. For example about the concept of " increased
internal body temperature " we need a function
providing meaningful coefficient indicating if
it true that it is increased and how far, regardless
of units (Fahrenheit, Celsius, ...), taking account of
meaningful physiologic values.
- Raw data available as classes should also be
standardized as a kind of coefficient. For example
pain scales from 1 to 10 and why not better from 0.0
to 1.0 .
- GWT, "Global Workspace Theory":
- Visual graph represent a small selection from the
global knowledge base. This could be seen as a kind of
awareness or consciousnesses at a given time.
- Relations between concepts:
- Identification:
- A set of relationships types between concepts will
be defined, for example "suggest", "confirm",
"exclude", "recommend", etc...
- Attributes of objects:
- Introduction:
- In general objects, both concepts and relationships,
have at least a kind of "weight" attribute, but many
more attributes should be considered.
- In many cases attributes values are optional and not
exclusive.
- Fuzzy Logic could require a standardization from 0.
to 1.
- Likelihood:
- Medical decision cannot be better than based on
likelhoods.
- Representation of likelihood in 3 different ways:
- Precise numerical figures intended to be used
inside the decision engine.
- Textual expressions intended for traditional
reports. Here the coefficients are rounded, for
example 0.9 as "evident", 0.7 as "probable", .5 as
"possible", 0.2 as "not to exclude". ... or in
time dimension 1 as "permanent", 0.7 as
"frequent", 0.3 as "occasional", 0.1 as "very
rare".
- Graphs visual attributes intended for quick
understanding of the situation, as for example
relative size of visual object in a graph, color
scale from red to blue.
- Specificity:
- Sensitivity:
- Positive or negative relation:
- Potential consequences, dangerousity ?
- Evolution, as acute or chronic.
- ...
- Synthesis of knowledge:
- We need synthesis made ready for decision support.
- Healthcare professionals in practice have not much
time nor interest to delve directly in the details of
scientific publications. Most of the time they relay on
opinions of trusted experts.
- A separate chapter is dealing with research issues,
about improvement of the current synthesis of medical
knowledge, particularly improvement of coefficients of
relationships.
- Procedures:
- Nodes may contain reference to procedures in the
decision support engine.
- Data Model:
- Most essential concepts:
- Knowledge base:
- Definition of 3 sets of concepts:
- Some "Observations", eg symptoms, lab
tests.
- Some "Problems", eg issues or diagnoses.
An abnormal observation may be also a problem.
- Some "Actions", seeking more information
or recommend treatment. For example which
questions should be asked, which lab test should
be ordered.
- Relations between these 3 sets:
- What most matter here is the relations between
concepts. Relations will be qualified with a
type, relative weights, etc...
- Here nodes represent knowledge, e.g. knowledge
abourt fever.
- Patients:
- A few experimental patient instances with:
- A few profile information, name sex, DoB, ..
- Having a few instance of results of
Observations.
- Will get a few instances of "Problems" with
their current relative likelihood.
- May get a list of recommended actions, sorted on
their relative priority. For example in emergency
situation what to do next.
- Care provider:
- ...
- Knowledge sources:
- Sources:
- Courses of professors teaching medicine in universities.
- CME, Continuing Medical Education networks, to be
converted and restructured as graphs.
- Agreed common opinions in scientific associations of
specialists, e.g. in cardiology, pneumology, gastro,
etc...
- Textbooks, if up to date.
- Many biomedical knowledge bass already exist:
- Traditional data bases:
- As far as possible extraction of knowledge by
means of NLP.
- Graphs:
- Some knowledge bases are already formatted as
graph but could need adaptation and extensions. For
example schemas of coagulation factors are already a
graph.
- Initial values of coefficients may need to be guessed
by experts and tuned later on.
- Multi-modal, accepting any kinds of information and
transformation to logical concepts.
- Anyhow medical experts have to validate the knowledge:
- A good human interface is here very important.
- Expected results:
- Make assisted Intelligence possible.
- Support of teaching and training of medical students.
- Work in progress:
- Results:
- Graph extracted from natural language sources are
already created from RxNorm,
MED-RT
and MeSH.
- A preliminary list of common relation types found in
these sources.
- Next steps:
- At a begin stage experimental work on relatively small
graphs in only a few limited medical domains. Every
contribution is welcomed even drawings on paper.
- More sources of knowledge must be explored as medicine
courses and textbooks.
- Seek more about the weights of relationships.
- Experimental setup of information about some specific
challenges, for example starting from a symptoms like
dyspnoea, icterus, antibiotic, ... with directly related
knowledge.
- Basic schemas examples:
- Starting from a given symptom, what could be the
Problem(s) ?
- Starting from every Problem, what are usual symptoms which
should be investigated ?
-
- In next versions attributes as likelihood, severity, etc
... will have visual styles like shape, size, color,
thickness, etc...
- Contact persons:
