The power of the

Cognitive Probability Graph(aka Cognitive Computing)

June 2016

Jans Aasman

ja@franz.com

10 years ago

Structured Data

7 years ago

Structured Data Unstructured Data

4 to 5 years ago

Structured DataUnstructured

Data

KnowledgeDomain knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontologies

New #1: Learning. Feed output of data

science back into data infrastructure

Structured

Data

Unstructured

Data

KnowledgeDomain knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontolo

gies

Probabilistic

Inferences.

New # 2: everything in one (distributed)

semantic graph

Structured

Data

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies

Probabilistic

Inferences.

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies

AKA: Cognitive Computing

Structured

Data

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies

Probabilistic

Inferences.

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies

Examples

Examples

• Healthcare: If I have this class of diagnostics and I get this procedure what are some of the new

symptoms I might get in the next two years.

• eCommerce and brand protection: find all my products based on product similarity

• Logistics: what can I statistically predict about part P breaking down and what other parts do I

usually buy after that part breaks down.

• Police Intelligence: find the most plausible story of a temporally orderend shortest path between

two criminal through observed (hard) facts and inferred (soft facts)

• Fraud detection: find links between your local chamber of commerce and the panama papers

through similar names and addresses.

Example healthcare

• Franz and Montefiore are partners in the Semantic Data Lake project.

One cognitive computing platform for all healthcare analytics

Example healthcare

• We created a single data centric platform that can serve any type of

analytic without building a new data mart for every new question.

• Currently 2.7 million patients with 10 years of data

• All data captured in a Unified Clinical Event model with 350 classes of

events.

Healthcare: structured and unstructured data

Structured patient data combined with complex

integrated terminology

Provenance for every value

Healthcare: the knowledge bases

• More > 180 vocabularies and terminology systems integrated in on

unified terminology system (Mesh, Snomed, UMLS, RxNorm, LOINC etc,

etc)

• External databases and

• Linked Open Data

OMOP

11089001

6600349

11894800

5

7534205

16790501

14667809

35896705

9209005

1732609

9908905

1469609

329005

LOINC

113345001 140460009

skos:semanticRelation

skos:narrower

118948005

skos:broader

“9209005”

SNOMEDCT

M0024135

M0008124 M0004742

skos:semanticRelation

skos:narrower

M0015742

skos:broader“Abdominal Pain”

“M0

02

41

35

”

skos:exactMatch

A0

54

93

02

A0978543

skos-xl:prefLabel

9209005

“Abdominal Pain”

SAB

AUI

SUI

MeSH

SNOMEDCT

MedDRA

rdfs:subClassOf

rdf:type

C0172359 C0232487

C0238551

“Abdominal Pain”

“C0000737”

skos:semanticRelation

skos:broader

UMLS - MTH

skos:notation

S035799

skos-xl:label

MTH

STR

C000737

Everything linked through SKOS

SK

OS

/SK

OS

-XL

ConceptScheme

Concept

UMLS - Semantic Net

Entity Event

Label

Population,

Community

Time

Pt.Pt.Pt.Pt.

SDL Paradigm:

Pt.Pt.Pt.Pt.

Diagnosis

Codes

Disease

Classification

OMIM,

GONG

Genetic

Profile

Procedure

CodeHCPC

Manufacturers

PharmKGB

Drug

Classification

Drug

Codes

DrugBank

ClinicalTrials

CER

PubMed

Analytic Tapestry(closed loop analytics)

Healthcare: probabilistic inferences

Why is this so important?

• Usually the output of data science results in reports and publications but

• No formal trace where the data came from

• No formal link to the actual methods you used, or who did it, or when you did it

• Cannot be compared to earlier results

• Cannot be used as building blocks for further research

• In general : the output is not queryable

• This is not good for delivery of care, reproducibility of research findings,

security and compliance, and results in loss of value-added information,

and enterprise intellectual property and assets, and unnecessary

duplication of efforts

Odds ratio

Association rules

K-means clustering

And then a query you could do never before

• Using the Knowledge Base, the Structured Data and the Probabilistic

inferences all at the same time.

• To find the statistical links between Diabetes and Vision problems in our

Semantic Data Lake

• Find the set of ICD9s that are connected via one or more steps to

concepts in the KB that mention Diabetes

• Find the set of ICD9s that are connected via one or more steps to

vision* or eye* or retinal*

• An show how those two sets are related in the space of odds ratios

And then just a few other examples

In the ecommerce world: find similar objects based on > 10

criteria, including description, product codes, pictures, etc

Returns a Graph in a Table (ughh )

But powerful when visualized

Or like this

And linking with the panama papers

And now the researchers can start investigating

Summary: this is the new paradigm of computing

Structured

Data

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies

Probabilistic

Inferences.

Unstructured

Data

KnowledgeDomain

knowledge

Linked Open Data

Vocabularies

Taxonomies/Ontol

ogies