Jalna Call Girls 👙 6297143586 👙 Genuine WhatsApp Number for Real Meet
Regeneron Dupixent Hackathon - D Cube Solution 20211217 v1.0
1. TM
D Cube Analytics
Pioneering a digitized analytics model to
optimize & transform life sciences
commercial operations
Proprietary and Confidential 1
2. TABLE OF CONTENTS
1. BACKGROUND & KEY BUSINESS QUESTIONS
2. BRAND CHALLENGES & PROPOSED SOLUTION
3. HACKATHON SOLUTION DEEP-DIVE
4. FUTURE ROADMAP
- Click on the corresponding boxes for navigating to their respective sections
Proprietary and Confidential 2
4. PROJECT BACKGROUND AND KEY BUSINESS QUESTIONS
4
Proprietary and Confidential
PROJECT BACKGROUND KEY BUSINESS QUESTIONS
• Dupixent is approved for three indications currently
and is expected to receive approval for new
indications in 2022
• Dupixent is prescribed widely by Allergists and PCPs
with the majority of prescriptions being driven by
Allergists
How do we drive incremental sales growth for the already
widely adopted Dupixent brand?
Can we better optimize call design considering the multiple
approved and upcoming indications for Dupixent?
6. BASED ON WHAT WE HEARD IN OUR DISCUSSIONS, WE UNDERSTAND THE DUP IXENT BRAND TEAM
HAS THREE KEY CHALLENGES TO SOLVE FOR
6
Proprietary and Confidential
Key Challenges Solution
• Dupixent is approved for three indications, and expected to
be approved for more indications
• Data-oriented basis needed to ascertain optimal call design
• While there is wide use of Dupixent, there continue to be
HCPs reluctant to expand Dupixent use to more indications
• Framework needed to encourage broader Dupixent use
amongst current writers
• Data across sales, patient claims, call and marketing channel
activity exists in a dis-jointed format
• Holistic approach to data aggregation and storage needed
Call Optimization Engine that optimizes number of indications
to detail, indication sequence, preferred channel (F2F vs virtual)
and messaging content
Cross-Sell Opportunity Identification to identify HCPs with
greatest potential to expand Dupixent use to new indications
Customer-360 data mart that hosts data across multiple data
sources in one location to act as the source of truth for analyze
needed across Dupixent functions
7. THE CUSTOMER 360 DATA MART WOULD FORM THE FOUNDATION POWERING AD VANCED
ANALYTICS APPLICATIONS TO ENABLE DUPIXENT GROWTH
Proprietary and Confidential
Customer 360
Data Model Creation
Call Hyper-
personalization
Cross-Sell Opportunity
Identification
Dynamic Targeting of HCPs
Next Best Action
Recommendation
Omni-Channel Marketing
Orchestration
Monitoring &
Measurement of Uptake
Scale to other Regeneron
brands
In Scope for Hackathon PoC
8. Go-To-Market AI
( GTM.AI )
OUR SOLUTION LEVERAGES ADVANCED MACHINE LEARNING TO OPTIMIZE TAR GETING AND CALL
PLANNING FOR INCREMENTAL DUPIXENT GROWTH
8
Proprietary and Confidential
• Cross-Sell Opportunity Prediction: Identify HCPs with “hidden”
opportunity for Dupixent sales growth, and encourage overall
growth via adoption of new indications
• Performance Dashboarding: Interactive Tableau Dashboard to
enable Dupixent stakeholders to drill down on HCP segments that
have Dupixent potential, and monitor performance KPIs
IDENTIFY OPPORTUNITIES
FOR INCREMENTAL
DUPIXENT SALES GROWTH
Objective Approach
Call Optimization
( Field.AI )
• AI/ML based Call Optimization to optimize for:
• Call Medium : In-person vs Virtual
• Number of Indications Detailed
• Priority Order of Indications Detailed
• Call Messaging Content
OPTIMIZE CALL PLAN DESIGN
TO DRIVE INCREMENTAL
DUPIXENT SALES GROWTH
9. Go-To-Market AI ( GTM.AI )
AI-powered module to enable nimble HCP targeting
Predictive Analytics to Unearth Latent
Dupixent Opportunity for Current Indications
Continuous Deployment on Dataiku to assess
multiple data sources and identify target HCPs
Growth Marketing Dashboard to enable
Dupixent stakeholders to drill down on
potential growth segments
Proprietary and Confidential 9
OUR SOLUTION WILL BE TAILORED TO THE SPECIFIC NEEDS OF THE DUPIX ENT BRAND TEAM, AND
INTEGRATED INTO REGENERON SYSTEMS TO ENSURE A SEAMLESS EXPERIENCE
Key Solution Tenets
Call Optimization ( Field.AI )
Deep Reinforcement Learning based Call Parameter Optimization
State Of The Deep Neural Networks based
Algorithm to Solve Problem Unique to Multi-
Indications brands like Dupixent
Scalable To Other Regeneron Brands with
Multiple Indications
Dynamic & Evolving Monthly Recommendations
For Most Optimal Call Type to replace existing
Static Call Recommendations
Seamless Integration with Veeva CRM to aid
sales reps, with feedback loop to improve model
10. EXECUTIVE SUMMARY
10
Proprietary and Confidential
• While ~86% of Allergists prescribe Dupixent, ~40% prescribe only for 1-2 indications
• Retrofit A/B testing shows that adoption of new indications results in a ~3% lift in
prescriptions for existing indications
• A cross-selling opportunity predictive model helps identify the next best HCP to target for
incremental Dupixent sales growth
Go-To-Market AI
( GTM.AI )
• Historical call activity and prescription data is leveraged to build a machine learning model
that derives affinity scores based on the likelihood of the call type resulting in a lift in
prescribing
• The ML model predicts the HCPs with an accuracy of 77%
• The model will be integrated with Veeva CRM to roll out optimal call design recommendations
on a monthly / quarterly basis
Call Optimization
( Field.AI )
12. DUPIXENT IS WIDELY ADOPTED, HOWEVER THERE IS OPPORTUNITY TO DRIV E DUPIXENT USE
AMONGST PRESCRIBERS WHO HAVEN’T YET ADOPTED ALL THREE INDICATION S
12
Proprietary and Confidential
14%
19%
21%
47%
% of HCPs
3 Indications
2 Indications
1 Indication
Dupixent Non-Writers
(n=3,416)
Distribution of HCPs Based on Indication Use
(Dupixent NBRx over past 12 months)
Does adopting a new indication result in incremental Dupixent writing for existing indications?
• There is wide adoption of Dupixent, with
~86% of Allergists having prescribed
Dupixent in the past 12 months
• Almost half of Allergists write Dupixent
across all 3 indications
• However, ~40% of prescribers continue
to limit Dupixent use to 1-2 indications
Insights
13. 13
Proprietary and Confidential
WE CONDUCTED AN A/B TEST TO UNDERSTAND IF A HALO EFFECT EXISTS DUE TO ADOPTION OF NEW DUPIXENT
INDICATIONS
A vs B Testing Approach
1000
1100
900
950
Prior 6M TRx Next 6M TRx
Cohort A Cohort B
+10%
+6%
Incremental lift of 4%
1. HCPs are divided into two cohorts based on Dupixent
prescription in a 6M vs 6M period:
• Cohort A: HCPs who began writing new indications
compared to the previous 6 months
• Cohort B: HCPs who continued writing the same
number of indications as the previous 6 months
2. The increase in Dupixent TRx is measured across two
groups, and incremental lift is measured
A vs B Testing Illustration
14. OVER THE PAST 12 MONTHS, ~39% OF HCPS ADOPTED A NEW DUPIXENT IND ICATION ; ADOPTING
NEW INDICATION(S) RESULTED IN ~3% LIFT IN PRESCRIBING FOR THE EX ISTING INDICATION(S)
14
Proprietary and Confidential
Adoption of New Indication(s)
(6 over 6 months)
Uptake in Prescriptions
(TRx per HCP)
24.5
27.7
0.0
10.0
20.0
30.0
Existing Indication
Previous 6M Next 6M
+13.0%
Continued
Same
Indication
(n=755)
Adopted
New
Indications
(n=483)
Incremental Lift of
+ 3.2 %
61%
39%
HCP (Allergists)
Adopted New Indication(s) Continued Same Indication
(n= 1238)
25.0
29.0
0.0
10.0
20.0
30.0
Existing Indication
Previous 6M Next 6M
+16.2%
15. FOR WRITERS WHO WERE PREVIOUSLY ONLY PRESCRIBING DUPIXENT FOR AS THMA, ADOPTION OF
NEW INDICATIONS RESULTED IN A ~6% LIFT IN ASTHMA WRITING
15
Proprietary and Confidential
Adoption of New Indication(s) Uptake in Prescriptions
(TRx per HCP)
7.0
0.0 0.0
7.3
0.0
2.0
4.0
6.0
8.0
10.0
AS AD NP
Pre Post
+4.6%
Continued
Same
Indication
(n=179)
6.7
0.0 0.0
7.4
11.1
3.4
0.0
5.0
10.0
15.0
AS AD NP
Pre Post
+10.3%
Adopted
New
Indications
(n=213)
Incremental Lift of
+ 5.6%
54%
46%
AS Only Writers
Adopted New Indication Continued Same Indication
(n= 392)
16. FOR WRITERS WHO WERE PREVIOUSLY ONLY PRESCRIBING DUPIXENT FOR AD , ADOPTION OF NEW
INDICATIONS RESULTED IN A ~9% LIFT IN AD WRITING
16
Proprietary and Confidential
0.0
17.5
0.0
0.0
20.4
0.0
0.0
10.0
20.0
30.0
40.0
AS AD NP
Pre Post
+16.6%
Continued
Same
Indication
(n=125)
0.0
22.7
0.0
4.2
28.5
2.7
0.0
10.0
20.0
30.0
40.0
AS AD NP
Pre Post
+25.6%
Adopted
New
Indications
(n=102)
Incremental Lift of
+ 9.0%
55%
45%
AD Only Writers
Adopted New Indication Continued Same Indication
Adoption of New Indication(s) Uptake in Prescriptions
(TRx per HCP)
(n= 227)
17. FOR WRITERS WHO WERE PREVIOUSLY ONLY PRESCRIBING DUPIXENT FOR AS THMA-AD, ADOPTION
OF NEW INDICATIONS RESULTED IN A ~1% LIFT IN ASTHMA-AD WRITING
17
Proprietary and Confidential
Adoption of New Indication(s) Uptake in Prescriptions
(TRx per HCP)
7.3
28.2
0.0
7.8
32.4
0.0
0.0
10.0
20.0
30.0
40.0
50.0
AS AD NP
Pre Post
+13.3%*
Continued
Same
Indication
(n=417)
7.7
34.6
0.0
8.3
40.2
6.8
0.0
10.0
20.0
30.0
40.0
50.0
AS AD NP
Pre Post
+14.5%*
Adopted
New
Indications
(n=202)
Incremental Lift of
+ 1.1%
67%
33%
AS & AD Writers
Adopted New Indication Continued Same Indication
(n= 619)
* : Cumulative lift across both AS and AD
18. D CUBE PROPOSES A COMPREHENSIVE GO -TO-MARKET MODULE (GTM.AI) TO IDENTIFY HCPS FOR
NEW INDICATION ADOPTION AND LEVERAGE HALO EFFECT TO DRIVE OVERAL L GROWTH
18
Proprietary and Confidential
Cross-Sell Opportunity Identification
• A Tableau dashboard will be developed to assess HCP
movement into new indications
• The dashboard will be enabled with dynamic cohort creation
and profiling, to enable Dupixent brand stakeholders to drill
down on demographics and behaviors of HCPs who stick to
one indication
Growth Marketing Dashboarding
• Hypothesize predictors of new indication adoption
• Derive features from raw data and validate hypotheses
Feature Engineering
• Iterate on predictive models to identify HCPs with greatest
likelihood of adopting new indication
• Finalize model based on performance metrics, for e.g. F-1
score
Predictive Modelling
• Deployment on Dataiku, and API-pathway based
integration with Veeva CRM
• Monthly/Quarterly refreshes to identify target HCPs
Deployment & Integration
19. THE ADOPTION OF NEW INDICATIONS IS DRIVEN BY MEASURABLE INDICATO RS OF POTENTIAL,
PRESCRIPTION BEHAVIORS, SAMPLE & CALL ACTIVITY
19
Proprietary and Confidential
Metric Dimension Importance Score *
Dupixent NRx share – Asthma (recent 6M) Asthma Prescribing Behaviors 12.7
Dupixent TRx share – AD (recent 12M) AD Prescribing Behaviors 12.6
Total Calls in Recent 6M Call Activity 12.0
Growth in AD Trx Volume (recent 12M) AD Biologic Potential 9.6
AD biologic TRx volume (recent 12M) AD Biologic Potential 5.8
Asthma biologic NRx volume (recent 6M) Asthma Biologic Potential 4.7
AD Eucrisa NRx Share (recent 6M) AD Prescribing Behaviors 3.6
Asthma NRx per patient (recent 12M) Treatment Proactivity 2.7
Total Staffed Beds Size of Practice 2.6
IL5 NRx Share (recent 6M) Affinity to IL5 Medications 2.2
HCP called upon in past 6M (1/0) Call Activity 1.5
Hcp received sample drop in past 6M (1/0) Sample Activity 1.0
* Note: The importance score is the mean decrease in accuracy when metric is removed from the model
Drivers of New Indication Adoption
Post the PoC, we will deploy a robust predictive model to identify HCPs with greatest potential for new indication adoption.
21. 21
Proprietary and Confidential
Process
Flow
DATA MODELLING PREDICTIVE MODELLING
DEPLOYMENT &
MONITORING
COHORT CREATION
DATA PREPARATION
DATA ENGINEERING
DATA MODEL
PRE-PROCESSING
MODEL TRAINING
BUSINESS RULE
IMPLEMENATION
A RIGOROUS ANALYTICS PROCESS WAS FOLLOWED TO DEVELOP THE FIELDAI MODEL FOR CALL
OPTIMIZATION
SCALING AND DEPLOYMENT OF THE SOLUTION ON REGENERON PLATFORMS WILL HAPPEN IN THE POST-POC PHASE.
Hackathon PoC Results
IMPROVE MODEL WITH
MORE ROBUST DATA
VEEVA CRM INTEGRATION
MODEL MONITORING
22. Proprietary and Confidential 22
A COMPREHENSIVE VIEW OF THE HCP WAS BUILT ACROSS MULTIPLE DIMENSIONS TO FEED THE MACHINE LEARNING
ALGORITHM
2
1
3
4
5
6
7 HCP 360 View
MCM- Response *
Response time stamp, Disposition code
Response/Contact, etc.
Medical Claims (Mx)
Trx, NRx, NBRx, Rx per patient, etc.
Call Activity
# of calls, sample drops
Prescriber Demographics
Specialty, Decile, etc.
Disposition *
Disposition code descriptions,
Engagement weights
Retail Claims (Rx)
Trx, NRx, NBRx, Rx per patient, etc.
Multi-Channel Marketing Activity *
# of Campaigns, Content
* Note: MCM data isn’t currently available and can be incorporated in the post POC phase
23. Proprietary and Confidential
DNN – Q learning, is a Deep Neural
Network with reinforcement learning,
which allows the algorithm to improve
the predictions with each cycle as it
various layers of artificial neural
network
The call optimization model will
optimize for three key parameters:
1. Call Channel – F2F vs Virtual
2. Number of Indications to Detail
3. Priority Order of Indications
4. Messaging Content*
The model will be triggered on a
monthly basis to generate optimal call
recommendations that will integrated
seamlessly with Veeva CRM via a
custom API.
Deep Neural Network – Q learning Optimized Call Output
Prescriber 1
F2F Call
Prescriber 2
23
Two Indications
AD -> AS
Efficacy *
Virtual
Call
One Indication
NP
Access *
NEURAL NETWORK MODELLING WAS LEVERAGED TO ASSESS MULTIPLE COMBINATIONS OF CALL PARAMETERS ALONG
WITH MARKET STATE METRICS TO BUILD A PREDICTIVE MODEL FOR NEXT BEST CALL TYPE
Veeva CRM Integration
Sigmoid
Data Science Engine Veeva CRM
Feedback loop for continuous learning
* Note: Messaging content data isn’t currently available and can be incorporated in the post POC phase
24. Proprietary and Confidential 24
THE NEURAL NETWORK MODEL PREDICTS THE PROPENSITY OF EACH CALL COMBINATION RESULTING IN AN INCREASE IN
PRESCRIPTION ACTIVITY
Model Setup
• For each call, we measure total prescriptions in the prior 3 months, and next 3 months
• Calls which results in >= 0.5% increase in prescriptions post call activity are labelled “Growth Call”, else “Regular
Connect”
Model Performance
Confusion Matrix
1,976 1,565
1,077 6,836
Actual
Predicted
Growth
Call
Regular
Connect
Growth Call Regular Connect
Precision of 65% & Recall of 56%
Accuracy of ~77%
• Accuracy of ~77%
• Precision of ~65% (Growth Call class)
• Recall of ~56% (Growth Call class)
• AUC score of 73%
25. HACKATHON DELIVERABLES
25
Proprietary and Confidential
S.No Document Description Link
1 Business Rule Document Contains the business rules being applied across all analyses
To be shared as part of final
deliverable
2 Code Runbook
Contains all the detailed steps needed to setup the environment to run the codes
and the sequence that needs to be followed to run the codes
To be shared as part of final
deliverable
3 Data Engineering Module
Contains scripts for each data source and describes the data transformation steps
required to bring the data in the modelling format
To be shared as part of final
deliverable
4 DNN-Q notebook
Contains scripts of Deep Neural Network – Q (DNN-Q) ran to predict the optimal
call type
To be shared as part of final
deliverable
5
Call Optimization Model
Results
Table with affinity score of each HCP to each call type
27. 01
02
03
04
05
06
Feature Engineering 2.0
✓ Integrate different data sources
✓ Develop features to be utilized for
reporting layer and data science
model
Call Optimizer Development
✓ Improve model performance
leveraging advanced AI/ML
technique
✓ Deploy model on Dataiku platform
GTM AI Model Development
✓ Improve model performance
leveraging advanced AI/ML
technique
✓ Deploy model on Dataiku platform
Scale to Other Brands
Growth Marketing Dashboard
✓ Build dashboard to monitor the
model performance and keep a
check on the engagement metrics
Standardize, Automate,
Productionize & Scale
✓ End to end automation &
standardization of data pipelines to
enable real time recommendations
27
Proprietary and Confidential
✓ Identify other brands with need for
similar solutions
✓ Design solution and deploy
NEXT STEPS TO TAKE THIS POC TO ROBUST, AUTOMATED & PRODUCTIONIZED RECOMMENDATION SYSTEM
28. Yes
✓ Single Database
capturing holistic
view of prescriber
✓ HCP related base
and derived
features
HOLISTIC FEATURE SET
MODEL TRAINING AND DEPLOYMENT
Weekly runs to generate
daily recommendations for
7 days from a previous
production model
Model Run
BUSINESS RULES
Yearly runs to Train Model
for different
hyperparameters using
Hyperopt and deploy Model
Retraining
PATH
-
3
Univariate
and
Bivariate
Analysis of
features
Final
Recommendations
Model
Performance
KPI changes
greater than a
threshold
value ?
KEY ENTITIES
PATH
-
1
Over-rides for
F2F/Virtual call
recommendations based
on accessibility
Filters for indication
sequence that business
team would not want to
roll out
Manual Intervention required
to make the decision of which
model to use after the analysis
No
Dataiku: Track, log, register, deploying and serving
models
Spark integration (SparkTrials) : Parallelize model
tuning in batches on multiple worker nodes
Hyperopt : General hyperparameter tuning library
for ML in Python
Custom
Models
Models Model Registry
Flavor 1
Flavor 2
Staging Production Archived
v1
v2
v3
General Model
Format that
standardizes
Deployment
options
Centralized and
collaborative model
lifecycle
management
Tracking Server
PATH
-
2
Incremental learning based
on additional data to
modify model weights, and
learn on the fly
Incremental Learning
THE MODEL WILL BE DEPLOYED ON DATAIKU AND OUTCOMES WILL BE INTEGRATED WITH VEEVA CRM FOR SALES FORCE
CONSUMPTION
29. T E N TAT I V E T I M E L I N E S F O R I M P L E M E N TAT I O N F O R D U P I X E N T B R A N D
Lorem ipsum
Feature
Engineering
Kick-off
GTM.AI
Development
Call Optimization
Development
Model Deployment
& Production
Wk1 Wk2 Wk3 Wk4 Wk5 Wk6 Wk7 Wk8 Wk9 WK10 WK11 Wk12 Wk13
Requirement
Finalization
29
Hyper-Care
30. WE BELIEVE IMPLEMENTING THE SOLUTIONS BASED ON THE HACKATHON POC WOULD TRULY ADD VALUE TO THE
DUPIXENT BRAND
State-of-the-art AI/ML
• The solutions will leverage advanced machine learning techniques, for e.g.
neural networks for call activity optimization
Dynamic Recommendations
• Dynamic models refreshed on a monthly / quarterly basis to recommend
next-best HCP targets, and optimal call design based on changing market
Automated Deployment &
Seamless Integration
• Solution deployed on Dataiku to enable “one-click” refreshes
• Call Design Recommendations integrated with Veeva to ensure excellent
experience for field force