Using Big Data for Improved Healthcare Operations and Analytics

Big Data for Healthcare:
Usage, Architecture and Technologies

Presenters

Pete Stiglich – Sr. Technical Architect
 Over 20 years IT experience

 Enterprise Data Architecture, Data Management, Data Modeling, Data Quality, DW/BI,
MDM, Metadata Management, Data Quality, Database Administration (DBA)

 President of DAMA Phoenix, writer, speaker, former editor Real World Decision Support,
listed expert for SearchDataManagement – Data Warehousing and Data Modeling

 Certified Data Management Professional (CDMP) and Certified Business Intelligence
Professional (CBIP), both at master level

Email: Pete.Stiglich@Perficient.com

Phone: 602-284-0992

Twitter: @pstiglich

Blog: http://blogs.perficient.com/healthcare/blog/author/pstiglich/

Presenters

Hari Rajagopal – Sr. Solution Architect
• Over 15 years IT experience

• SOA solutions, Enterprise Service Bus technologies, Data Architecture, Algorithms

• Presenter at conferences, Author and Blogger

• IBM certified SOA solutions designer

Email: Hari.Rajagopal@Perficient.com

Phone: 303-517-9634

Key Takeaway Points

• Big Data technologies represent a major paradigm shift – and is
here to stay!

• Big Data enables “all” the data to be leveraged for new insight–
clinical notes, medical literature, OR videos, X-rays, consultation
recordings, streaming medical device data, etc.

• More intelligent enterprise – more efficient and prevalent
advanced analytics (predictive data mining, text mining, etc.)

• Big Data will affect application development and data
management

Agenda

• What is Big Data?

 How Big Data can enable better healthcare

 Types of Big Data processing

 Key technologies

 Impacts of Big Data on:

 Application Development

 Data Management

 Q&A

What is “Big Data”

• Datasets which are too large, grow too rapidly, or are too
varied to handle using traditional techniques

• Volume, Velocity, Variety

• Volume – 100’s of TB’s, petabytes, and beyond

• Velocity – e.g., machine generated data, medical devices,
sensors

• Variety – unstructured data, many formats, varying
semantics

• Not every data problem is a “Big Data” problem!!

MPP enables Big Data

100’s, 1,000’s of nodes

Scalability Scalability

Cluster (homogenous) or Grid (heterogeneous)

SMP – Symmetric MPP – Massively Parallel
Multiprocessing Processing
“Shared Everything” “Shared Nothing”
CPU, memory, disk (SAN, NAS) Nodes do not share
CPU, memory, disk (DAS)

Cost Factor

 Cost of storing and analyzing Big Data can be driven down
by:

 Low cost commodity hardware

 Open source software

 Public Cloud? Yes, But for really massive amounts of data with many
accesses, may be cost prohibitive

 Learning curve? You bet!

Hadoop / MapReduce

• Hadoop and MapReduce – key Big Data technologies
developed at Google, now open source

• “Divide and conquer” approach

• Highly fault tolerant – nodes are expected to fail

• Every data block (by default) replicated on 3 nodes
(is also rack aware)

• MapReduce – component of Hadoop, programming
framework for distributed processing

• Not the only Big Data technology…

NoSQL

• Stands for “Not only SQL” – really s/b “Not only Relational”

 New(ish) paradigms for storing and retrieving data

 Many Big Data platforms don’t use a RDBMS

 Might take too long to setup / change

 Problems with certain types of queries (e.g., social media, ragged
hierarchies)

 Key Types of NoSQL Data Stores
• Key-Value Pair
• Wide Column
• Graph
• Document
• Object
• XML

How can “Big Data” improve Healthcare?

Healthcare “Big Data” opportunities

• Examples of Big Data opportunities
 Patient Monitoring – inpatient, ICU, ER, home health

 Personalized Medicine

 Population health management / ACO

 Epidemiology

 Keeping abreast of medical literature

 Research

 Many more…


• Patient Monitoring

 Big Data can enable Complex Event Processing (CEP) – dealing with
multiple, large streams of data in real-time from medical devices,
sensors, RFID, etc.

 Proactively address risk, improve quality, improve processes, etc.

 Data might not be persisted – Big Data can be used for distributed
processing with the data located only in memory

 Example – an HL7 A01 message (admit a patient) received for an
inpatient visit – but no PV1 Assigned Patient Location received within X
hours. Is the patient on a gurney in a hallway somewhere???

 Example – home health sensor in a bed indicates patient hasn’t gotten
out of bed for X number of hours


• Personalized Medicine
 Genomic, proteomic, and metabolic data is large, complex, and varied

 Can have gigabytes of data for a single patient

 Use case examples - protein footprints, gene expression

 Difficult to use with a relational database, XML performance problematic

 Use wide-column stores, graphs, key-value stores (or combinations) for better
scalability and performance

Source:
wikipedia


• Population Management
 Preventative care for ACO – micro-segmentation of patients

 Identify most at risk patients – allocate resources wisely to help these
patients (e.g., 1% of 100,000 patients had 30% of the costs)*

 Reduce admits/re-admits, ER visits, etc.

 Identify potential causes for infections, readmissions (e.g., which two
materials when used together are correlated with high rates of infection)

 Even with structured data, data mining can be time consuming – distributed
processing can speed up data mining

* http://nyr.kr/L8o1Ag (New
Yorker article)


• Epidemiology
 Analysis of patterns and trends in health issues across a geography

 Tracking of the spread of disease based on streaming data

 Visualization of global outbreaks enabling the determination of ‘source’ of infection

17


• Unstructured data analysis
 Most data (80%) resides in unstructured or semi-structured sources – and a wealth
of information might be gleaned

 One company allows dermatology patients to upload pictures on a regular basis to
analyze moles in an automated fashion to check for melanoma based on redness,
asymmetry, thickness, etc.

 A lot of information contained in clinical notes, but hard to extract

 Providers can’t keep abreast of medical literature – even specialists! Use Big Data
and Semantic Web technologies to identify highly relevant literature

 Sentiment analysis – using surveys, social media

 Etc…

Poll

• What Healthcare Big Data use case do you see as being most
important for your organization?

• Patient Monitoring
• Personalized Medicine
• Population Management (e.g., for ACO)
• Epidemiology
• More effective use of medical literature
• Medical research
• Unstructured data analysis
• Quality Improvement
• Other

19

Analytics

• Big Data ideal for experimental / discovery analytics

• Faster setup, data quality not as critical

• Enables Data Scientists to formulate and investigate
hypotheses more rapidly, with less expense

• May discover useful knowledge . . . or not

• Fail faster – so as to move on to the next hypothesis !

Unstructured Data Mining

• Big Data can make mining unstructured sources(text, audio,
video, image) more prevalent - more cost effective, with better
performance

• E.g., extract structured information, categorize documents,
analyze shapes, coloration, how long was a video viewed, etc.

• Text Mining capabilities
• Entity Extraction – extracting names, locations, dates, products, diseases, Rx,
conditions, etc., from text

• Topic Tracking – track information of interest to a user

• Categorization – categorize a document based on wordcounts/synonyms, etc.

• Clustering – grouping similar documents

• Concept Linking – related documents based on shared concepts

• Question Answering – try to find best answer based on user’s environment

Data Mining

Text
• Can enable much faster data mining

• Can bypass some setup and modeling Text Mining
effort
Other use Entity
cases Extraction
• Data Mining is “the automatic or semi-automatic
analysis of large quantities of data to extract
previously unknown interesting patterns” Wikipedia Data
Structured
Data Mining
• Examples of data mining:

• Association analysis - e.g., which 2 or 3 Something
materials when used together are correlated Interesting?
with a high degree of infection

• Cluster analysis – e.g., patient micro-
segmentation

• Anomaly / Outlier Detection –e.g., network
breaches

Transaction Processing

• Some Big Data platforms can be used for some types of
transaction processing

• Where performance is more important than consistency e.g.,
a Facebook user updating his/her status

• More on this later…

Poll

• What type of Big Data use case would be most beneficial for
your client?

• Complex Event Processing (using massive/numerous
streams of real-time data)

• Unstructured Data Analysis

• Predictive Data Mining

• Transaction Processing (where performance more
important than consistency)

25

Big Data Architecture and Key Technologies

Hadoop

• Used for batch processing – inserts/appends only – no updates

• Single master – works across many nodes, but only a single data
center

• Key components

• HDFS – Hadoop Distributed File System

• MapReduce – Distributes data in key value pairs across nodes, parallel
processing, summarize results

• Hbase – database built on top of Hadoop (with interactive capabilities)

• Hive – SQL like query tool (converts to MapReduce)

• Pig – Higher level execution language (vs. having to use Java, Python) –
converts to MapRduce

28

Cassandra

• Used for real-time processing / transaction processing

• Multiple masters – works across many nodes and many data
centers

• Key components

• CFS – Cassandra File Systems

• CQL – Cassandra Query Language (SQL like)

• Tunable consistency for writes or reads. E.g., option to ensure a write
succeeds to each replica in all data centers before returning control to
program …. or can be much less restrictive

29

In memory processing

• To support real-time operations, an IMDB (In-Memory Database)
may be used

• Solo – or in conjunction with a disk based DBMS

• I/O most expensive part of computing – using in memory database /cache
reduces bottlenecks

• Can be distributed (e.g., memcache, Terracotta, Kx)

• Relational or non-relational

• E.g., for a DW, current values might reside in an IMDB, historical data on disk

30

MPP RDBMS

• Have been in around for 15+ years

• Used for large scale Data Warehousing

• Ideal where lots of joins are needed on massive amount of data

• Many NoSQL DB’s rely on 100% denormalization. Many do not
support join operations (e.g., wide column stores) or updates

31

Semantic Web

• Semantic Web – web of data, not documents

• Machine learning (inferencing) can be enabled via Semantic Web
technologies. May use a graph database/triplestore (e.g.,
Neo4j, Allegrograph, Meronymy)

• Bridge the semantic divide (varying vocabularies) with
ontologies – helps address the “Variety” aspect of Big Data

• Encapsulate data values, metadata, joins, logic, business rules,
ontologies, access methods in the data via common logical model
(e.g., RDF triples) – very powerful for automation, federated
queries

32

Semantic Web
Find Jane Doe’s relatives (with machine inferencing)

System X System Y System Z

a:JoeDoe :isInLaw

:hasBrother :hasBrother

:marriedTo
x:DebDoe y:JohnDoe z:JaneDoe
:hasBrother

:isInLaw
Original data
Inferred data

33

No One Size Fits All

 Many types of solutions will require multiple data
paradigms

 E.g. Facebook uses MySQL (relational), Hadoop, Cassandra,
Hive, etc., for the different types of processing required

 Be sure to have a solid use case before deciding to use Big
Data / NoSQL technology

 Provide solid business and technical justification

What type of data store to use??

Big Data impact on Application Development
and Data Management

ACID / CAP / BASE

 If your transaction processing application must be ACID compliant, you must
use an RDBMS (or ODBMS)

 ACID – Atomic, Consistent, Isolated, Durable

Atomic – All tasks in a transaction succeed – or none do
Consistent – Adheres to db rules, no partially completed transactions
Isolated – Transactions can’t see data from other uncommitted transactions
Durable – Committed transaction persists even if system fails

 Not all transactions require ACID – eventual consistency may be adequate

Vs..

ACID / CAP / BASE

 Brewer’s CAP theorum for distributed database

 Consistency, Availability, Partition Tolerance - Pick 2!

 For Big Data, BASE is alternative for ACID

Basically Available – data will be available for requests, might not be consistent

Soft state – due to eventual consistency, the system might be continually changing

Eventually consistent – the system will eventually be consistent when input stops

• Example: HBase every transaction will execute, but only the most recent for a
key will persist (LILO – last in, last out) – no locking

Data Management

 Security not as mature with NoSQL – might use OS level encryption (e.g.,, IBM
Guardium Encryption Expert, Gazzanga) - encyrpt/decrypt at IO level

 Data Governance needs to oversee Big Data – new knowledge uncovered can
lead to risks - privacy, intellectual property, regulatory compliance, etc.

• Physical Data Modeling less important – due to “schema-less” nature of NoSQL

• Conceptual Modeling still important for understanding business objects and
relationships
• Semantic modeling – inform ontologies which enable inferencing
• Logical Data Modeling still useful for reasoning and communicating about how
data will be organized

• Due to schema-less nature of NoSQL – metadata management will be more
important!
• E.g., wide-column store with billions of records and millions of variable columns
– useless unless you have the metadata to understand the data

Getting started

• Data Scientist is a key role in Big Data – requires statistics, data modeling, and
programming skills. Not many around and expect to pay $$$’s.

• Big Data technologies represent a significant paradigm shift. Be sure to allow budget
for training, sandbox environment, etc.

• Start small with Big Data . Start with a single use case – allocate significant
amount of time for learning curve, and environment setup, testing, tuning,
management.

• Working with open source software can present challenges. Investigate purchase of
value added software for simplification. Tools such as IBM Big Insights, EMC
Greenplum UAP (Unified Analytics Platform) adds analytical, administration, workflow,
security, and other functionality.

40

Summary

 Big Data presents significant opportunities

 Big Data is distinguished by volume, velocity, and variety

 Big Data is not just Hadoop / MapReduce and not just NoSQL

 Key enabler for Big Data is Massively Parallel Processing (MPP)

 Using commodity hardware and open source software are options to drive
down cost of Big Data

 Big Data and NoSQL technologies require a learning curve, and will continue to
mature

Resources

 Perficient Healthcare: http://healthcare.perficient.com

 Perficient Healthcare IT blog: http://blogs.perficient.com/healthcare/

 Perficient Healthcare Twitter: @Perficient_HC

 Apache – download and learn more about Hadoop, Cassandra, etc.

 http://hadoop.apache.org/

 http://cassandra.apache.org/

 Comprehensive list with description of NoSQL databases: http://nosql-
database.org/links.html

 Translational Medicine Ontology (TMO) - applying Semantic Web for
personalized medicine: http://www.w3.org/wiki/HCLSIG/PharmaOntology

About Perficient

Perficient is a leading information technology consulting firm serving
clients throughout North America.

We help clients implement business-driven technology solutions that
integrate business processes, improve worker productivity, increase
customer loyalty and create a more agile enterprise to better respond
to new business opportunities.

PRFT Profile
 Founded in 1997

 Public, NASDAQ: PRFT

 2011 Revenue of $260 million

 20 major market locations throughout North America
— Atlanta, Austin, Charlotte, Chicago, Cincinnati, Cleveland,
Columbus, Dallas, Denver, Detroit, Fairfax, Houston,
Indianapolis, Minneapolis, New Orleans, Philadelphia, San
Francisco, San Jose, St. Louis and Toronto

 1,800+ colleagues

 Dedicated solution practices

 600+ enterprise clients (2011) and 85% repeat business
rate

 Alliance partnerships with major technology vendors

 Multiple vendor/industry technology and growth awards

Our Solutions Expertise & Services

Business-Driven Solutions Perficient Services
• Enterprise Portals  End-to-End Solution Delivery
• SOA and Business Process  IT Strategic Consulting
Management  IT Architecture Planning
• Business Intelligence  Business Process & Workflow
• User-Centered Custom Applications Consulting
• CRM Solutions  Usability and UI Consulting
• Enterprise Performance Management  Custom Application Development
• Customer Self-Service  Offshore Development
• eCommerce & Product Information  Package Selection, Implementation
Management and Integration
• Enterprise Content Management  Architecture & Application Migrations
• Industry-Specific Solutions  Education
• Mobile Technology
• Security Assessments
Perficient brings deep solutions expertise and offers
a complete set of flexible services to help clients
implement business-driven IT solutions 47

Using Big Data for Improved Healthcare Operations and Analytics

In this document