1.
Knowledge Graph based
segmentation of HCP with
heterogenous dimension data
for pharma commercial
acceleration
Sridhar Nomula

2.
Background
Chronic diseases are slow progression and are the result of a combination of
genetic, physiological, environmental and behaviors factors.
• Approx. Every 40 seconds, someone in the USA will have a myocardial infarction.
• On average in 2019, someone died of stroke every 3 minutes 30 seconds
7.6M
0
2000000
4000000
6000000
8000000
10000000
12000000
2 0 1 2 2 0 1 5 2 0 2 0 2 0 2 5 2 0 3 0 2 0 3 5 2 0 4 0
HP
POPULATION
P R OJECTIONS O F T H E U S P O PULATION WI TH
H E ART F A ILURE F R OM 2010 T O 2040
Prevention, control and early detection of cardiovascular risk for
individual patients are vital interventions.
Every patient requires individual and special treatment. Identifying the right
patient for the pharma CVD product is critical factor for to increase the
volume & quality of HCP.
No. of Pharmaceutical CVD products have been put on the market over
the past decade, and 200-plus drugs are in clinical trials, according to
PharMA

3.
Research Question
Largest segment of a
market is the best and
most profitable to
pursue. However, if all
products in the market
are chasing the largest
segment of that market,
then this segment
consequentially
becomes the most
competitive.
Why the easy way of
targeting HCPs doesn’t
always work is that
although an HCP may
prescribe in high volume,
that HCP may not be
amenable to prescribing
our product. Ex., An HCP
may be so loyal to a
competitive product that
the amount of sales effort
sales rep would have to
expend to convert the
HCP would make that
HCP no longer profitable.
Targeting HCPs is not
right way in most
circumstances.
We predict the chances of a
developing serious heart failures
early in its onset?
Encoding the domain
the form rules will improves the
patient and physician
accuracy?

4.
4
Challenges/Risks
Low quality of high-
volume dataset
Diagnostic data
misses the big picture
Rx data are summary
numbers
Diagnostic information
is too complex to
actionize.
Data is siloed and
departmentalized
Knowledge graph
• High dimensionality causes the model instability due to high parameters learning.
heterogeneous data sources integration is missing.
• Explainability is the need of hour business to thrust the model predictions

5.
Data
Extraction
Ontology
Entity
relationship
ETL Database
Visualization
Reasoning
engine.
Query
example
things
Graph
problem
formulation
Training and
test split
GNN model
building
Calculate
the loss
metric
Write back
to Graph DB
This approach has two parts
 Knowledge graph building
 Machine learning model building
Methodology
Knowledge Graph
Samples
Rx-claim
Calls Dx-claims
Affiliatio
ns
Procedur
e claim
Market
access
Drugs
Insurance
LL
Data
Sources
Graphs are a general language
for describing and analyzing
entities with
relations/interactions
Patient
HCP Rep
Payer
Sample node edge Graph
Enables to capture,
organize, & query a large
amount of multi-relational
data and making it possible
to infer new knowledge by
reasoning engine.

6.
Ontology- The structure for our data
Knowledge model gives Comprehensive patient view and Physician view

7.
Reasoning Engine
• Infer new facts from the existing
graph such as patient journey
• Explainability and able to validate
the hypnosis made.
• Reasoning is lacking in majority of
the systems in today's world
define
rule medical-trans:
when {
$p isa patient;
$d1 isa drug;
$d2 isa drug;
$d3 isa drug;
$d4 isa drug;
$d5 isa drug;
$d6 isa drug;
$d7 isa drug;
$r1(reimbursement-for:$p, $d1) isa rx-claim;
$r2(reimbursement-for:$p, $d2) isa rx-claim;
$r3(reimbursement-for:$p, $d2) isa rx-claim;
$r4(reimbursement-for:$p, $d2) isa rx-claim;
$r5(reimbursement-for:$p, $d2) isa rx-claim;
$r6(reimbursement-for:$p, $d2) isa rx-claim;
$r7(reimbursement-for:$p, $d2) isa rx-claim;
$r1 has fill-date $dt1;
$r2 has fill-date $dt2;
$r3 has fill-date $dt3;
$r4 has fill-date $dt4;
$r5 has fill-date $dt5;
$r6 has fill-date $dt6;
$r7 has fill-date $dt7;
$dt1 < $dt2;
$dt2 < $dt3;
$dt3 < $dt4;
$dt4 < $dt5;
$dt5 < $dt6;
$dt6 < $dt7;
} then {
(victim:$p, prog1:$d1,prog2:$d2, prog3:$d3, prog4:$d4, prog5:$d5, prog6:$d6, prog7:$d7) isa
disease-progression;
};
Patient Brand A
Brand B
Brand C
Brand D
Brand A

8.
 Given a node, examine the nodes in the vicinity of that node.
 Heterogeneous graph neural network can model complex system; Use different
network weights for different relation types.
 Learns meta paths automatically, handles heterogeneity
8
Deep learning on Graph
Embeddings via supervised learning
Neighborhood aggregation Message passing Algorithm Multi relational graph algorithm
Query
example
things
Graph
problem
formula
tion
Training
and test
split
GNN
model
building
Calculat
e the
loss
metric
Write
back to
Graph
DB

9.
Results
Test accuracy ● The selected data sample size is
of 1815,790 patients, and the
median age was 69 years.
● learning rate = 0.005, weight
decay = 0.001, Adam optimizer,
type encoding size = 16 and
attribute encoding size = 16.
● HGT model outperforms the HAN
model and HEAT (Edge
enhanced Graph attention
Network)
● Recall 0.94 and F1 score 0.87
Loss metric and accuracy
Test metrics and accuracy
Model Loss Precision Recall F1 score
HEAT 0.41 0.70 0.88 0.77
HAN 0.34 0.82 0.89 0.84
HGT 0.32 0.81 0.94 0.87
Models' metrics comparison

10.
System Design
Structured Semi-
structured
Un-
structured
RDBMS,
No-SQL
etc.,
XML, Email,
etc.,
Documents
, posts etc.,
Knowledge
graph
ML
Reports
query
ETL
User interface
Data
Information
Knowledge
deductive
logic
Data that are placed in a meaningful and useful context for a
user; are manipulated, presented, and interpreted using
ontology
A knowledge graph gets richer as new data is
added. Data representation of facts, observation, and
occurrences.
Knowledge is the application of data and information; answers
“how” questions.

11.
 Graph database: TypeDB (2.10.2)
 Package Manager: Anaconda. Python 3.8
 VS code with TypeQL extension
 Type Studio windows (2.5.0)
 Data processing: pandas, NumPy, json
 Data manipulation with TypeDB: typedb-client
 Modeling: pytorch, sklearn, PyG, typedb-ml
 Visualization: NetworkX, pytorch-tensorboard
Software & Hardware
System Requirements
Operating System windows/2016
RAM 64 GB
CPU Memory 16GB
vCPUs 8 CPUs
AWS Instance Type p3.2xlarge
Number of Nodes 1

12.
Benefits

13.
Conclusion
• ML algorithms can perform better if they can incorporate domain
knowledge.
• Iterative deep learning-based knowledge representation technique is
much more accurate than other conventional machine learning.
• Experimented with other state of the art algorithms and observed
2.1% increment in overall accuracy HGT.
• Due to the computation challenges, the experiment is performed on
the small sample data.
• Predictions can be explainable and detailed factor analysis of
predictions is not carried out in the experimental setup.

14.
Contributions
 This study identified that it is possible to make use
of heterogonous data that bring context to the data.
 Experimented with the help of Graph software,
Python and built the link prediction model that
gives a probability of patient being critical and
physician likely to reach.
 This is the first study to pharma commercial
acceleration with advanced methodology powered
by AI

15.
Future works
• Not all treatments affect every patient in the same way, is
something doctors are aware of. Yet for decades, the strategy
of trial and error has still been used to a large extent to treat
and diagnose patients.
• Personalized medicine is an evolving field in which physicians
use diagnostic tests to identify specific biological markers,
often genetic, that help determine which medical treatments
and procedures will work best for each patient.
• Diagnostic testing plays a vital role in the precision medicine
approach for selection of appropriate treatment based on
patients’ genetic makeup.
• Diagnosing complex diseases like cancer can be a time and
labor/Lab-intensive process. Proposed Methodology will
improve the personalized care of serious patients with explain
ability
The global precision medicine market size was valued at USD 66.1 billion in
2021 and is expected to reach USD 175.64 billion by 2030
Concert Genetics says that 39 new genetic tests enter the market
every day
13.8%
PD growth
forecast