REAL-WORLD BIG DATA ANALYTICS

FOR UNDERSTANDING TREATMENT

PATHWAYS

John Cai

Director, Medical Informatics

June 11, 2015

Big Data for Pharma

Decision

making

The Fourth Hurdle requires Real

World Evidence from RWD

Cost-effectiveness (or CER) has became

the “fourth hurdle” to market access

Real World Big Data

Complexity

Variety Unstructured

data types

e.g. clinical

notes

Volume Massive data

sets, e.g.

longitudinal

claims/EMR

Velocity Fast, real-time

data collection

and

transmission

e.g. HIE,

wearables

Volume: Real World Population and

Real World Data

• Real World Evidence (RWE) evaluates safety, effectiveness and outcomes using real

world data (RWD).

• Not RCT data and broader than observational data, RWD is health data collected from

actual practice by healthcare providers or in day-to-day situations by patients or

caregivers

Real World

Population

Randomized

Clinical Trial

Population

100

1,000

10,000

100,000

1,000,000

10,000,000

Phase 1 Phase 2 Phase 3 Phase 4 5 yrs 10 yrs

Typical Pharma Data

Real World Data

#patients

Observational

Study Population

Variety: Major Real World Data

Types and Sources

• Claims (from payers or data vendors): Truven (MarketScan), IMS (PharMetrics),

United Health Group (Optum), Wellpoint, Aetna, Humana, CMS, ...

• EMR/EHR (from Healthcare providers or EMR vendors):

Nation-wide: VA, DoD, GE Centricity, Allscripts, Cerner, Humedica, Flatiron, etc…

Regional: Kaiser, Regenstrief, Partners, Mayo, Intermountain, Geisinger, ...

Academic: Harvard, Univ of Utah, Vanderbilt, Cincinnati Children's Hospital, ...

• Surveys and registries: NCHS (NHANES, NHIS, NAMCS , NHAMCS, NSAS,

NHDS, NNHS, NNAS, etc.), SEER registries, MEPS, ACC registries, ...

• PBM/Pharmacy Databases: Medco, Wallgreens, CVS, Walmart, …

• Lab databases: Quest, Labcorp, …

• PHRs: patient portals, MS HealthVault™, Indivo X, CMS PHR Pilots, …

• Patient forums/social media: Patientslikeme, inspire.com, smartpatients.com…

• Monitoring/wearables: medical device data, Apple ResearchKit, …

Pharma

CER EBM

Proactive

Pharmacovigilance Trial Design

& Interpretation PHC

Cost

Effectiveness

Drug Repositioning

/New Indications

Patient

recruitment

Velocity: Real World Data

Transmission to Pharma

Payer/

PBM Real World Data

?

?

Complexity, Variability, Veracity

• Patient journeys are complex

• Real-world treatment

pathways can be messy

• Physicians not following

clinical practice

guidelines

• Patients not adherence

to medications

Treatment pathways are difficult to

reconstruct using healthcare data:

• Technical hurdles - need to repeatedly

query and merge across large # tables

• Conceptual hurdles of secondary use

• Claims for transaction

• EMR for patient care

9

• Use business rules to translate data to events of interest

- Example: ndMM patient cohort

One inpatient diagnosis or two outpatient diagnoses (two separate dates)

list of ICD9 codes

One or more MM-specific treatments

list of drugs and procedures

First diagnosis: “index date”

At least 6 or 12 months continuous coverage before index date

At least 12 or 24 months continuous coverage after index date

What is a therapy line?

What is a drug switch, discontinuation, add-on, combo, “drug holiday”?

• Addresses some parts of the conceptual challenge

• Creates new problems

- How sensitive are our results to the rule definitions?

Typical Solutions

Potential Technical Solution:

Hadoop and MapReduce

• Hadoop: an open source software project

- Hadoop Distributed File System (HDFS)

- MapReduce: compute paradigm for parallel computing

- A whole ecosystem of additional products/services/tools

• History:

- 2003 Google file system paper

- 2004 Google Map Reduce paper

- Adopted by Yahoo, donated to the open source community in 2009

• The gist of it:

- Distributed file system, “cheap” storage on computer clusters

- Compute paradigm that abstracts the parallelism by breaking down

operations to “map” and “reduce”

- Hadoop framework takes care of everything else

Map Reduce in a Nutshell

Mappers work on data,

“emit” key-value pairs

We write Mappers and Reducers

Hadoop takes care of everything else

Reducer works on all

values (data) for the

same key

Shuffle-Sort:

intermediary data

sorted and distributed

by key

12

• Load data into HDFS

- “Transactional” data (claims, interactions)

• Reconstructing a patient’s timeline is a textbook MapReduce exercise:

- Mapper:

Read a piece of data. Example: claim

Figure out who it relates to. Example: patient ID

Return key-value pairs:

Key: patient ID

Value: the full piece of information (claim)

- Reducer:

Gets as an input a key and the set of all values (claims) associated with that key (patient ID)

Organize the values (claims) to produce a basic patient history

Building Patient Timelines using Hadoop and MapReduce

13

Building Patient Timelines using MapReduce Followed by Visual Analytics

Shuffle-Sort: “Hadoop magic”

Mapper Reducer

Treatment Cost Trends

14

Cost analysis of PsA and PsO treatments

Biologics treatment costs have been high and going up

Presented to AMCP and ISPOR 2015 as posters

Co-medication Usage

Treatment Pathways

Patient timelines - “individual story”

Future Directions

Cost of care analysis, comparing across different pathways

Healthcare resource utilization analysis, comparing across different pathways

Patterns of care analysis: predictive modeling combining patient similarity measures and clustering

Comparison to Clinical Practice Guidelines (Compliance and Adherence)

Outcomes of care/CER: incorporating clinical outcomes using integrated claims/EMR data

Some Learning Points

Some Hadoop functionality perfectly suited for patient timeline analysis Mapreduce for creating patient timelines

Once patient timelines are created, everything else scales linearly

Map(reduce) for calculating patient metrics and complex events

Mapreduce for analyzing treatment pathways

Cheap scalable storage capacity and compute power Scalability allows robust analysis

Healthcare Decision Making Requires

Real-world Big Data Analytics

Efficacy and Safety from RCT settings – FDA to approve

Cost effectiveness – Payer's willingness to pay

Clinical effectiveness (long term efficacy and safety) – Physicians to

prescribe, patient to adhere

Comparative effectiveness, patient reported outcomes – Physicians to

prescribe, patient to adhere

To Innovate To Approve To Pay for To Prescribe To Adhere

Industry FDA Physician Patient

Health Plan

IDS

Government

Forthcoming

Thank You!

Leveraging Hadoop MapReduce in Building Patient Timelines and Analyzing

Health Resource Utilization

Special Issue on Big Data in Pharmacoeconomics

Saar Golde, Ph.D., Knowledgent Group and NYU

Zhaohui “John” Cai, M.D. Ph.D., Celgene Corporation