can we fight cancer with big data? - hisa · can we fight cancer with big data? august 7, 2017 ken...

© Flatiron Health 2017

Can We Fight Cancer

With Big Data?

August 7, 2017

Ken Carson, MD, PhD

Senior Medical Director, Flatiron Health

Assistant Professor of Medicine,

Washington University School of Medicine

The Emerging Role of Real World Evidence (RWE)

1


YES!2


Case Study 1:

PDL1 in NSCLC Cohort

3


Cohort DemographicsAs of July 31, 2017

4

Histology Smoking Status

Patients in cohort: 35,512 (Community: 32,032 | Academic: 3,480)

History of

smoking

No history of

smoking

Unknown / not

documentedNot otherwise

specified

Non-squamous

cell carcinoma

Squamous cell

carcinoma


PDL1 Biomarker Test OverviewTesting rate among actively treated patients, as of July 31, 2017

5

© Flatiron Health 2017 6

Patient Share by Therapy Class — PD1/PDL1All Lines

10% 40%


Patient Share by Therapy Class — PD1/PDL11st Line

4%

32%


Patient Share by Therapy Class — PD1/PDL12nd or 3rd Line+

20% 51%


Where are we now?

9

Role of real-world evidence in oncology drug

development


Real-World Data

10

● Information derived from health care delivered outside the traditional clinical trial

setting

● Includes EHRs, claims/admin data, registries, and patient-generated health data

● Complements clinical trials – more generalizable knowledge

The current deficit in evidence has become

particularly acute for the FDA, which in numerous

areas lacks vital evidence needed to support definitive

regulatory determinations.”– Robert Califf

Former Commissioner, FDA


Clinical

Practice

Retrospective Capture with Longitudinal Follow-Up

Consent with or without

randomization

Prospective Capture

Observational data on

off label use for

regulatory purposes

Real-world follow-up on

clinical outcomesPragmatic trial

Identification of rare

patients

Real-World

Data

Clinical Trial

Dataset

Retrospective RWD Prospective RWD

Continuum of Real-World Evidence


What are the data

requirements for

regulatory grade

RWE?

● Aggregated, high quality,

complete, longitudinal

● Reproducibility and provenance

● Patient-level data linkage

● Endpoints/outcomes

● A priori identification of study

objectives and analysis plans

● Careful cohort selection


Key steps in preparing EHR data to

develop real-world evidence

13

© F️latiron Health 2017

Develop a world-

class team:

We come from

places like...

14


2M

Active Patients

2,500

Clinicians

265

Cancer Clinics

800

Unique Sites of Care

The Flatiron Network

Aggregate EHR into a single common dataset


Demographics

Diagnosis Visits

Labs Therapies

Discharge NotesPathology

Physician Notes

Radiology Report

EHR

Hospital Reports

RWE

Database

Structured

Data

Processing

Unstructured

Data

Processing

16

Standardize EHR data to a common data model


Standardize EHR data to a common data model

1751-7Albumin [Mass/volume]

in Serum or Plasmag/dL

● Certain structured data elements may be

coded and collected in multiple ways in the

EHR across practices (example: albumin)

● Combine and map datasets across sites to a

single dataset

● Map all data elements to a single set of

definitions (data model)

2220 Blood Serum Albumin g/dL

QD25001600 ALBUMIN/GLOBULIN RATIO QD (calc)

QD25001400 ALBUMIN QD g/dL

QD50058600 ALBUMIN %

QD50055700 ALBUMIN g/dL

CL3215104 Albumin % (EPR) %

LC001081 ALBUMIN, SERUM (001081) g/dL

LC003718 Albumin, U %

LC001488 Albumin g/dL


CL3215162 Albumin%, Urine %

CL3215160 Albumin, Urine mg/24hr

3234 ALBUMIN SS g/dL


QD50060710 MICROALBUMIN mg/dL

QD50061100MICROALBUMIN/CREATININE RATIO,

RANDOM URINEmcg/mg creat

QD85991610 ALBUMIN relative %

50058600 ALBUMIN UPEP RAND %

CL3210074 ALBUMIN LEVEL g/dL

QD86008211 ALBUMIN/GLOBULIN RATIO (calc)

LC149520 Albumin g/dL

QD45069600 PREALBUMIN mg/dL

QD900415245 ALBUMIN, SERUM mg/dl

QD900429745 ALBUMIN g/dL

CL3215124 Albumin Electrophoresis g/dL

LC016931 Prealbumin mg/dL

QD50060800 MICROALBUMIN, 24 HOUR UR mg/24 h

QD50060900 MICROALBUMIN, 24 HOUR UR mcg/min

QD85994821 ALBUMIN,SERUM g/dL

CL3213320 PREALBUMIN mg/dL

QD85995225 PROTEIN ELECTROPHORESIS ALBUMIN g/dL

Harmonization and normalization of

structured data


Key data elements in oncology are unstructured

For every PD-1/PD-L1 test a patient

receives, Flatiron biomarker Data Model

captures:

● Test result

● Date biopsy collected

● Date biopsy received by laboratory

● Date result received by provider

● Lab name

● Sample type

● Tissue collection site

● Type of test (e.g., IHC)

● Assay / kit (e.g., Dako 22C3)

● Percent staining & staining intensity

Result

Result

Lab Name

Tissue Collection Site

Section of PD-L1 Report


Standardize EHR data to a common data model:Curate unstructured data

● Abstraction of unstructured data through Patient Manager allows for

comparability across patients as well as different oncology practices

Technology-Enabled Abstraction using Patient Manager®

EHR Source Data

Pathology report

Radiology report

Physician notes


Resulting clinical data is research-grade

VariableStructured data

only

Flatiron data

completeness

Metastatic

diagnosis26% 100%

Smoking status 0%1 94%

Histology 37% 99%2

Stage 61% 95%

ALK results

(of those tested)9% 100%3

EGFR results

(of those tested)11% 99%3

1 58% are free text in dedicated field in EHR (requiring hand abstraction)

2 Including 8% of patients with results pending or indeterminate test

3 Including 6% of patients with results pending or unsuccessful test

Completeness of technology-enabled

abstractionExample: Advanced NSCLC

Site of metInter-abstractor

agreementKappa

Bone 97% 0.93

Brain 96% 0.91

Liver 92% 0.83

Lung 94% 0.87

Accuracy of technology-enabled

abstraction

Example: Sites of metastases


Oncology research needs linked, annotated data

Germline & somatic

genomic data Claims data

BiosensorsProspective data

capture

High quality structured &

unstructured data, combined

and ready for research

Electronic health

records

Patient-reported

outcomes


Documentation of source, quality and

provenance.

Patient Stage at Dx Biomarkers 2L Treatment Progression Date of Death

Jane Doe II EGFR-, ALK-, PD-L1- nivolumab 2017-03-08 2017-04-12

Diagnosed with Stage

II NSCLC

Undergoes surgery for early-stage disease

Develops metastatic

disease

Tested for EGFR and ALK

Progresses on 1L, tested for PD-L1

and re-tested for EGFR

DeathStarts 1L therapy

Starts 2L therapy, deteriorates

and is hospitalized


Documentation of source, quality and

provenance.

Patient Stage at Dx Biomarkers 2L Treatment Progression Date of Death

Jane Doe II EGFR-, ALK-, PD-L1- nivolumab 2017-03-08 2017-04-12

Diagnosed with Stage

II NSCLC

Undergoes surgery for early-stage disease

Develops metastatic

disease

Tested for EGFR and ALK

Progresses on 1L, tested for PD-L1

and re-tested for EGFR

DeathStarts 1L therapy

Starts 2L therapy, deteriorates

and is hospitalized

> Abstracted by Sue Smith on 4/30/17 at 10:10am

> Physician notes and scan interpretation reviewed

> Medical record from West Florida Cancer Clinic

Quality of Progression abstraction

===================================

> Completeness: 99%

> Sue Smith is 96% accurate at last testing

> Inter-abstractor agreement: 97%

> Kappa: 0.93

> Audit trail for any changes

> Dataset freeze and storage

Abstraction Details


Case Study 2:

Patients excluded from a clinical trial

24

Evaluating patients who receive a drug with a boxed

warning


Trastuzumab

emtansine usage

in low-LVEF

patientsRWE Case Study

Overview:

● Anti-HER2 agents have known

cardiotoxicity

● Major Kadcyla clinical studies

excluded patients at risk for

cardiotoxicity: EF <50% or h/o

CHF, serious cardiac arrhythmia

requiring treatment, MI or

unstable angina w/i 6 mo

25


Objectives:

● To assess the rate and severity

of cardiac events in patients with

metastatic breast cancer treated

with trastuzumab emtansine who

have a LVEF ≤50% at treatment

initiation

● Characterize patient population

● Other outcomes (OS, PFS,

resource utilization)

Trastuzumab

emtansine usage

in low-LVEF



Study Design: Eligibility Criteria

● Diagnosis of breast cancer (ICD-9 174.x or

ICD-10 C50.x)

● At least two visits in the Flatiron Health

database on or after 1/1/2013

● Pathology consistent with breast cancer

● Evidence of stage IV or recurrent

metastatic breast cancer with a metastatic

diagnosis date on or after 1/1/2013

● Treatment with trastuzumab-emtansine

(structured med admin data and confirmed

via unstructured data review)

● LVEF ≤50% at the time of initiation of

treatment with trastuzumab-emtansine

Trastuzumab

emtansine usage

in low-LVEF



Cohort Selection

● Maintain a growing curated

dataset in mBC

● At time of cohort selection

>8,000 cases

● Allowed for the

identification of >60

patients who had LVEF

≤50%

From structured data, patients with:

ICD 9/10 diagnosis codes for breast cancer

Two visits on or after 1/1/2013

Structured medication order or administration for trastuzumab-emtansine

N = 1,149

From unstructured data review, confirmation of diagnosis with metastatic

breast cancer on or after 1/1/2013

N = 1,070

From unstructured data review, confirmation of treatment with

trastuzumab-emtansine

N = 1,000

From unstructured data review, LVEF ≤50% at the time of trastuzumab-

emtansine initiation, as defined by the most recent LVEF assessment

prior to trastuzumab-emtansine initiation

N = 66


● Complex information in

chart

● Duplicate abstraction

● Clinical adjudication

Data Quality Control

Assessing data quality of cardiac

information


RWE Case Study: LVEF Data (N=66)

Breakdown By LVEF Value Breakdown By Time Before Treatment


RWE Case Study: Patient Journey

31



32

Kadcyla

LVEF 45%



33

Monitor LVEF over time

Hospitalizations:

Discern if cardiac or cancer related


Case Study 3:

Observational data to complement

trials

34

Use of observational data to examine effectiveness of

approved agents used in the off-label setting


RWE Case Study:Observational data for off-label use of approved agents

● BRAF V600 inhibitors approved in melanoma

● Previously conducted Phase II basket trial in non-melanoma cancers with signal of

response in non-small cell lung cancer (N=20)

● Among people with BRAF+ non-small cell lung cancer, is there differential improved

response when a BRAF inhibitor is administered?

● Because of the difficulty of conducting clinical trials in this small population, sponsor

asked if it is possible to use real world data to supplement the clinical trials data

Context

● Size of database (>25,000 aNSCLC patients) allowed for the identification of >35

patients who had the BRAF V600 mutation, tested as part of routine clinical care

● Centralized, technology-enabled processing allowed for systematic assessment of

outcomes by capturing “real-world” tumor endpoints, based on standardized

methodology

RWE Application


History of NGS

testing

Structured order

for a

BRAF inhibitor

Free-text search

for

BRAF mutation

BRAF V600E

mutated

N = xxx

Treated with a

BRAF inhibitor

N = x

Not treated with a

BRAF inhibitor

N = x

Confirmed to meet

NSCLC

classification

criteria*

N = 27,729

27,729

>1000

<50

RWE Case Study:


Observations to Date From Oncology

Evolving portfolio of example studies where EHR-generated RWD has been

transformed into RWE used in regulatory and clinical decision-making

Data quality must be characterized and optimized

Approaches incorporate the longitudinal nature of clinical care

Use appropriate analysis plans and technology tools to support efficient

understanding of information

37

can we fight cancer with big data? - hisa · can we fight cancer with big data? august 7, 2017 ken...

Documents