Danbury AI

Danbury AI

What do we mean by “AI”?

- AI is an umbrella term for many different subfields which are grouped together by the concept of making programs that behave “intelligently.”

One way of looking at AI:

Meetup Goals

Bring people together around a common interest for mutual enrichment.

- Share AI information, news, and opportunities.
- Have monthly talks on interesting AI topics.
- Develop a team for Kaggle,etc competitions.
- Create a local hub for AI discussion, research, and collaboration.
- Create a pool of computational resources for research.

Member Archetypes

We are a free meetup open to the public. The key feature we wish members to have is a passion for learning or mastering AI topics.

You can be:

- A student.
- An expert.
- A novice.
- Anyone!

We want to provide resources and systems that allow everyone to flourish!

Introductions

Let’s take a moment to get acquainted. Please tell us the following:

- Your name
- Education, Skill Background, or Profession
- What drew you to the group? What about AI interests you?
- What would you like to get out of an AI meetup?

Introductions after we give them some context.

Meetup Structure

- Monthly meetings with interesting talks.
- We need speakers! Please let us know if you are interested in giving a talk!
- First Tuesday of every month. Starts at 7pm.
- Online Collaboration and Connection
- Slack, Medium Publication, Meetup & Forums, Linkedin Group, Facebook Group.
- News letter. ( possible )
- Weekly aggregate of AI news and DanburyAI community communiqués.
- Suggestions? Message us!

Key Links & Resources

- Github: https://github.com/DanburyAI
- LinkedIn Group: https://www.linkedin.com/groups/8537797
- Facebook Group: https://www.facebook.com/DanburyAI/
- Meetup: http://www.meetup.com/DanburyAI/
- Medium Publication: https://medium.com/danbury-ai
- Slack: Ask for an invite!

Our speaker:

Brian Griner, PhD

Quick Facts

- Ph.D., Public Policy Research & Analysis from University of Pittsburgh.

Mention that we hold questions till the end..

Using Bayesian Networks for Segmentation, Targeting & Position of New Products

Bayesian Networks: Nothing to it, right?

The Computer-based Patient Case Study (CPCS) Bayesian Network^{*} ^{*}(Pradhan et al. 1994) – 422 nodes: 14 disease, 33 history & risk, 375 health outcomes

A Definition

Bayesian Network [bey-zee-uhn net-wurk]

1. Informal: A series of interconnected models represented by a specific type of graph (i.e., a Directed Acyclical Graph or DAG)

Example of a Bayesian network with variables as “nodes” and relationships as “arcs”

Physician Characteristics

Patient Characteristics

Age

Severity

Specialty

Current

Rx

Attitude

3

Attitude

1

Attitude

2

Attitude

segment

Symptom

3

Symptom

1

Symptom

2

Disease

Activity

- Nodes represent variables as distributions
- Arcs represent relationships between variables

Behind each node is a conditional probability model linking the node to each of its ‘parent’ nodes

There are multiple relationships being described here.

The distribution of a node (or variable) depends on the distribution of its “parent” nodes (i.e., the nodes that are connected to the target node by arcs)

Let’s pause for a minute and let me show you a very simple example of that relationship.

How a node in a Bayesian network operates

X_{1}_{}

X_{2}_{}

0.1

0.3

0.6

1,1

1,0

0,1

0.4

0.2

0.4

0,0

0.5

0.3

0.2

lo

med

hi

Y

1

0

Parent Nodes (X_{1},X_{2})

Child Node (Y)

lo : 0.7

med : 0.1

hi : 0.2

Prediction (y | x_{1},x_{2})

(x_{1},x_{2})

0,1

0.7

0.1

0.2

0.7

0.1

0.2

0.7

0.1

0.2

Make just 3 main points here.

- Conditional probability model is very similar to a regression model
- We have a dependent variable (Child Node, Y) and independent variables (Parent Nodes, X
_{1}and X_{2}) - The predicted distribution of the Child Node depends on the distributions of the Parent Nodes

Comparison

| Regression Model | Bayesian Network Model |

Model Framework | One model | Series of models, |

Dependent, Independent Variables | One dependent variable | Every variable can be |

Hypothesis Testing | Test relationships between each independent variable and the dependent variable | Test relationship between |

Prediction | Predicts average value of the dependent variable given the levels of each independent variable | Model based on joint probability of Predicts the relationships of all the variables in a network simultaneously Predicts the entire joint distribution, not just one dependent variable |

Benefits:

- Link individual models into a system
- Measure uncertainty in the relationships between variables
- X -> Y becomes prob(X) -> prob(Y|X)
- Can easily incorporate beliefs and assumptions
- E.g. connecting nodes with known causal relationships
- Predict the change in one or more of the variables on the entire system

Case Study

Simulating treatment flows and opportunity for

newly launched asthma product

Approach

Bayesian network model provided an intuitive graphical model that:

- Simulated treatment decisions across physician and patient types
- Identified targeted patient and physician segmentation
- Sized the market opportunity
- Provided promotional material guidance

- New drug
- Challenging profile
- Crowded market

- Multiple data inputs
- EMR, Primary Research

and Publications - Weighted segment sizes

and treatment flows

to the population

Business Situation

Business Need

- Treatment flow
- Stakeholder segmentation
- Market sizing
- Promotional strategy

Need:

- A new drug launched is a crowded market (asthma)
- Concern with side effects (steroids stunt growth in children)
- Client needed to identify patient and physician segments of early adopters for targeted promotions

Approach:

- Data inputs: EMR, primary case-based research and published research
- Model: Bayesian network model created using conditional distributions from EMR and primary research.
- Weights: Published research and secondary data used to create weights to project sizes of patient and physician segments and treatment flows to the population

Results:

- Bayesian network model used to identify early adopter patient and physician segments
- Weighted model used to size the market opportunity and target key physician segments with patient promotional materials

Bayesian Network Case Study

New product launch in crowded Asthma market

Bayesian Network of Asthma Patient Treatment Flow

Select baseline for comparison

Step 1: Select all patients that were NOT candidates for the new therapy

Select target patient type for new therapy

Step 2: Select all patients that were candidates for new therapy

Step 3: Identify patient and physician characteristics of target segments

More characteristics of target segments

Segment profiles provide rich customer portraits of the target physicians and the patients who would be early adopters of the new therapy

In Conclusion

- Bayesian Networks provide a simple way to link individual models into a system of cause and effect relationships
- BNs can predict the impact of a change in one or more variables on ALL variables across the network
- BNs are useful tools for simulating market dynamics (e.g., buying process) and identifying joint physician and patient segments for targeted promotions