In the second of our Real World Data (RWD) webinars, we examined new techniques that go beyond the standard Body Mass Index, and how large data sets are being mined for meaningful real world applications.
Speakers included:
Dr. Naomi Allen, Senior epidemiologist, UK Biobank
Olof Dahlqvist Leinhard, PhD, Co-Founder & Chief Technology Officer, AMRA
Theresa Tuthill, PhD, Head of Imaging Methodologies, Biomarkers and Development Group, Early Clinical Development, Pfizer.
3. ŠPistoiaAlliance
The Panel
3
Olof Dahlqvist Leinhard, PhD, Co-Founder & Chief Scientific Officer, AMRA
Olof Dahlqvist Leinhard is AMRA's Chief Scientific Officer and Co-Founder, with background as an MR
physicist. He is responsible for AMRAâs technical vision, for leading the execution of technology
platforms, for overseeing technology research and product development, and for aiding in the clinical
translation of AMRA's research findings. He also teaches and runs a research group at the Centre for
Medical Image Science and Visualisation (CMIV) at LinkĂśping University, Sweden.
Dr. Naomi Allen, BSc MSc Dphil Senior epidemiologist, UK Biobank
Naomi Allen is an Associate Professor in Epidemiology and Senior Epidemiologist for UK Biobank. She
isresponsible for processing the linkage of routine electronic medical records into the study for long-term
follow-up (including deaths, cancers, primary and secondary care data as well as other health-related
datasets). She helps to co-ordinate the introduction of new enhancements into the resource (such as the
development of web- based questionnaires and proposals for cohort-wide biomarker assays) and
provides scientific advice to researchers worldwide wishing to access UK Biobank.
october 25, 2017 Beyond BMI - Body Composition Phenotyping in the UK Biobank
Theresa Tuthill, PhD, Head of Imaging Methodologies, Biomarkers and Development
Group, Early Clinical Development, Pfizer
Theresa Tuthill, PhD, is Head of the Imaging Methodologies, Biomarkers and Development group within
Early Clinical Development at Pfizer. Though trained as an Electrical Engineer, she oversees a small
group dedicated to the development of imaging biomarkers for metabolic, cardiovascular, and safety
applications in clinical trials.
4. Poll Question 1: Are you currently using UK biobank data?
A. Yes, I personally do
B. No, but my organization does
C. No, but I/we plan to in the future
D. No
5. Improving the health of future generations
www.ukbiobank.ac.uk
Overview of UK Biobank
Naomi Allen
naomi.allen@ndph.ox.ac.uk
6. UK Biobank is a major national health resource
designed to improve the prevention, diagnosis and
treatment of a wide range of illnesses that affect middle
and older age
Aim of UK Biobank
7. UK Biobank in a nutshell
⢠A large prospective cohort study
⢠500,000 UK adults age 40-69 at
recruitment, 2006-2010
⢠Baseline data on a wide range of
lifestyle factors, environment,
medical history, physical
measures & biological samples
⢠Consent for follow-up through
health records for all types of
health research
⢠Open-access to researchers
worldwide (academia & industry)
8. Recruitment into UK Biobank
⢠Using individual GP
practices for
recruitment
purposes impractical
⢠Direct mailing of
invitations using
contact details held
by the NHS
⢠Invited 9.2 million;
5.5% response rate
10. ⢠Socio-demographic information
⢠Lifestyle factors (diet, physical activity,
smoking, sleep)
⢠Environmental exposures
⢠Reproductive history & screening
⢠Sexual history
⢠Family history of common diseases
⢠General health & medical history
Large subsets
⢠Noise exposure
⢠Psychological status
⢠Cognitive function tests
⢠Hearing test
⢠Blood pressure
⢠Hand grip strength
⢠Body composition
⢠Lung function test
⢠Heel ultrasound
Large subsets
⢠Vascular reactivity
⢠Exercise test/ECG
⢠Eye measures (visual acuity,
refractive error, OCT scan)
Touchscreen questions Physical measures
Baseline assessment
11. ⢠Blood
⢠Whole blood
⢠Serum
⢠Plasma
⢠Red blood cells
⢠Buffy coat
⢠Urine
⢠Saliva
Total: 15 million 0.85ml
aliquots
Biological samples collected
13. ⢠Genotyping: Bespoke Affymetrix array
of 850,000 genome-wide genetic
markers
⢠Imputation: ~90 million genetic variants
⢠Data for all 500,000 participants made
available July 2017
⢠Largest study in the world with
genotyping, lifestyle and imaging data
⢠Exome-wide sequencing: Initiative
between UK Biobank and
Regeneron/GSK for all 500,000
participants
Genetic analysis of samples
14. ⢠Aim: to perform multi-modal imaging scans on 100,000
participants, 2014-2023
⢠Brain, cardiac and whole body MRI, carotid ultrasound and
whole-body DXA scans
⢠Can define phenotypes closely related to disease and
investigate how genetics and lifestyle factors influence
intermediate precursors of disease
Imaging: heart, brain, bones and body
15. ⢠Over 16,000 people have already been scanned
⢠Imaging centres in Stockport, Newcastle (Reading to
be opened March-April 2018)
⢠Opportunities for repeat imaging in 10,000
⢠Biggest study of its kind ever undertaken
⢠Collaboration with academic and commercial partners
to generate imaging derived phenotypes
UK Biobank Imaging Study
16. Death notifications: 14,000 participants
Cancer registrations: 79,000 participants
Hospital admissions: 400,000 participants
Primary care records: 230,000 so far
⢠to be made available 2018
Linkages to electronic health records
17. Access to UK Biobank
⢠Opened for access March 2012
⢠Available to all bona fide researchers
â Academic and commercial
â UK and international
⢠5,700 approved registrations
⢠1,000 applications submitted
â 700 projects approved and underway
⢠250 publications
⢠Apply online at www.ukbiobank.ac.uk
18. Poll Question 2: Are you using imaging biomarkers?
A. Yes, I personally do
B. No, but my organization does
C. No, but I/we plan to in the future
D. No
19. The Body Composition Profile
Enhancing the Understanding of Metabolic
Syndrome using UK Biobank Imaging Data
Olof Dahlqvist Leinhard, MSc, PhD
Advanced MR Analytics AB, AMRA, LinkĂśping, Sweden
Center for Medical Image Science and Visualization, CMIV
LinkĂśping University, LinkĂśping, Sweden
CENTER FOR MEDICAL IMAGE
SCIENCE AND VISUALIZATION, CMIV
olof.dahlqvist.leinhard@liu.se
Chief Scientific Officer, Founder
20. From Population Medicine to Precision Medicine
6.8 L5.2 L0.7 L 1.6 L 2.2 L 3.2 L
Different Body Compositions. Different Metabolic Risk.
Visceral
Adipose
Tissue
Six Men with BMI 21
21. AMRAÂŽ Profiler Research
A New Standard in Body Composition
Rapid
6-Minute
MRI
4 Individualized
3
Platform Agnostic
Modern 1.5 and 3T
GE, Siemens and Philips
2 Accurate & Precise
1 3D Volumetric
23. Cancer
Yesterday and Todayâs Approach to Cancer
Today
Cancer Research UK; http://www.cancerresearchuk.org/about-cancer/what-is-cancer.
Yesterday
200 types of cancers & treatments
31. Comparison to Dallas Heart Study (DHS) Results
⢠VAT was quantified in 973 obese subjects and followed for 9.1 years
⢠Doubled risk for CVD events in high VAT subjects
32.
33. Health Care Burden
⢠Based on Health Episode Statistics (HES) Data
⢠From United Kingdomâs secondary care hospital
services
⢠Collected to allow hospitals to be paid for
delivered care
⢠Includes, e.g., information of diagnosis and
operations, and administration
⢠Definition: Number of hospital nights
truncated at 30 nights
⢠Standardized way of reporting
⢠Requires referral by physician
⢠Robust to outliers
⢠Insensitive to type and amount of ICD-10 codes
Frequency
Nbr of nights hospitalization
34. Statistical modelling
BCP Effect on Health Care Burden
VATi ASATi Liver Fat IMAT
Univariate
p-value
*** *** *** ***
đ˝-value 0.34 Âą 0.04 0.21 Âą 0.03 0.23 Âą 0.07 0.15 Âą 0.02
Multivariate
p-value
*** n.s.
** ***
đ˝-value 0.30 Âą 0.07 - â0.29 Âą 0.09 0.09 Âą 0.02
* p < 0.05, ** p < 0.01, *** p < 0.001, n.s. non-significant
BCP Effect on Health Care Burden
Statistical results adjusted for sex and age
1. West J. ECO Annual Congress 2017: Oral presentation OS7:OC65.
2. Romu T. ECO Annual Congress 2017: Poster T1P59.
35.
36. LinkĂśping University
⢠Anette Karlsson
⢠Thord Andersson
⢠Per Widholm
⢠Thobias Romu
AMRA
⢠Jennifer Linge
⢠Janne West
⢠Patrik Tunon
⢠Brandon Whitcher
⢠Magnus Borga
Pfizer
⢠Theresa Tuthill
⢠Melissa Miller
⢠Alexandra Dumitriu
Acknowledgement
University of Westminster
⢠Jimmy Bell
⢠Louise Thomas
Imperial College
⢠Alexandra Blakemore
⢠Andrianos Yiorkas
This research has been
conducted using the UK
Biobank Resource.
(Access application 6569)
CENTER FOR MEDICAL IMAGE
SCIENCE AND VISUALIZATION, CMIV
38. Theresa Tuthill, PhD
Imaging, Pfizer
Radiomics for Metabolic Disease:
Mining Large Data Sets
Pistoia Alliance
October 25, 2017
39. Radiomics
⢠Radiomics â defined as the
conversion of images to higher
dimensional data and the
subsequent mining of these data
for improved decision support.
⢠Also known as ⌠Imiomics
⢠The mining of radiomic data to
detect correlations with genomic
patterns is known as
radiogenomics.
⢠Most commonly used in Oncology
to characterize tumors.
Gillies RJ, et al. Radiology 2015;278:563â77.
Aerts, HJWL, et al. Nature communications 5 (2014).
Coroller, TP, et al. Radiotherapy and Oncology 119.3 (2016): 480-486
40. Oncology Example
Used to discriminate between cancers that progress quickly and those that are stable.
⢠Patterns of change can be predictive of response to treatment.
⢠Early studies showed a relationship between quantitative image features and gene expression
patterns in patients with cancer
Gillies RJ, et al. Radiology 2015;278:563â77.
Include tumor texture, blood flow,
cell density, necrosis, etc
41. Challenges with Imaging Biomarkers
⢠Distinction between imaging biomarkers
and bio-specimen derived biomarkers.
â Scanners are designed to produce images which are interpreted by diagnostic radiologists
â Innovation is largely driven by competition to improve image quality
â Quantified measurements are often vendor-specific
⢠Key Issues for Imaging Biomarkers
â Validation of technology
⢠Repeatability/reproducibility
â Need for standardization of acquisition
â Data reduction
⢠Whole body scan can contain millions of measurements
â Clinical Use : Diagnostic and/or treatment?
vs
42. Radiomic Analysis for Understanding Disease
⢠Creating predictive models involves receiving input
from clinical data, radiology data, pathology data,
protein data and gene testing data
â Larger data sets provide more power
⢠Look at imaging data and the various â-omicâ data
(radiomics, pathomics, proteomics, genomics) to
discover their relationship with each other
⢠A multidisciplinary data-mining effort involving
radiologists, medical physicists, statisticians, bio-
informatists, geneticists, and other researchers
â Imaging parameters need standardized acquisition
and analysis (segmentation, regions of interest, etc.)
Clinical
Data
Pathology
Data
Radiology
Data
Gillies RJ, et al. Radiology 2015;278:563â77.dat
43. Characterizing Body Types with Disease Risk
Current standard is to use BMI and Waist Hip Ratio
Visceral obesity:
Increased risk of
macrovascular disease
Peripheral obesity:
Decreased risk of
metabolic disease
Fu, J et al. Cell metabolism 21.4 (2015): 507-508.
Lebovitz, HE, International journal of clinical practice. Supplement 134 (2003): 18-27.
44. Alternative Body Composition: Need standardization
⢠VAT and SAT can be estimated from CT and MR
images
â Single slice imaging poorly predicts VAT and SAT
changes in longitudinal studies1
⢠Whole body MRI allows more complete estimation
⢠AMRA has standardized quantification2
â Automated segmentation of fat and muscle
â VAT defined as the adipose tissue within the
abdominal cavity
â ASAT defined as subcutaneous adipose tissue in
the abdomen from the top of the femoral head to
the top of the thoracic vertebrae T9
1Shen, W., et al. Obesity 20.12 (2012): 2458-2463.
2West, J., et al. PloS one 11.9 (2016): e0163332.
45. Large Imaging Databanks to Mine?
⢠UK Biobank â Started in 2006
â 500,000 subjects in age range 40 - 69 years
â Collected measures included blood, urine and saliva samples (genome-
wide genetic data and biomarker panel available on all subjects)
â Access to electronic medical records
â Imaging subcohort â 7,000 in Pilot Project, May 2014 - October 2015
⢠Single imaging site in Stockport, NW England
⢠3 adjacent imaging suites:
â MRI (Brain, full body & heart),
â DXA (Bone density)
â Carotid ultrasound
⢠AMRA body composition analysis of full body MRI
http://www.ukbiobank.ac.uk/
46. Defining Disease Groups
⢠Use hospital in-patient records
â Filter based on ICD-10 codes
⢠Activity based on questionnaire
⢠For healthy cohort, remove subjects with âŚ
â Cardiovascular disease
â Metabolic disease
â Cancer, strong infectious diseases, etc.
47. Healthy women have lower liver fat and VAT than healthy men.
0 5 10 15 20
0
10
20
30
40
50
60
VAT
Frequency
Male
Female
0 5 10 15 20
0
10
20
30
40
50
60
Liver Fat Fraction (%)
Frequency
Male
Female
95th Percentile
Liver Fat VAT
Female 3.8 2.9
Male 6.0 4.6
49. Clustering by Characteristics to Find Natural Groupings
Need algorithms for higher dimensional data
What features should be used?
Should the data be normalized?
Does the data contain any outliers?
Jain, AK. Pattern recognition letters 31.8 (2010): 651-666
50. Unsupervised Clustering of Body Composition Profile
High
Low
Color Key and Histogram
Male Female Together
51. Phenomapping through Cluster Analysis
⢠Clustering based on body composition
parameters
⢠Identify subgroups that may underlie
metabolically un-healthy subjects
⢠Define and characterize mutually exclusive
groups
â Blinded to disease outcomes
⢠Within a cluster, determine the number of
subjects with a specific self-reported
disease
⢠Compare this ratio with that of the
combined remaining clusters
⢠Ultimate goal is to define therapeutically
homogeneous patient subclasses
56. Target ATarget B
Target C
Target ATarget B
Target C
Radiomics to Inform Clinical Trials
⢠What is cutoff for âHealthy liver fatâ?
â For patient identification
⢠What is the distribution of liver fat in selected cohorts?
â For inclusion/exclusion criteria
⢠What genetic loci are associated with liver fat?
â For target identification and target validation
⢠What are phenotypic clusterings?
â For patient stratification Match pathway intervention to
patientâs pathogenic trajectory
57. Using Data to Aid in Patient Stratification for Clinical Trials
0 5 10 15 20 25 30 35
0
5
10
15
20
25
30
General Population
Liver Fat Fraction (%)
Frequency
Controls
Diabetics
0 5 10 15 20 25 30 35
0
5
10
15
20
25
30
Population w/ BMI>28
Liver Fat Fraction (%)
Frequency
Controls
Diabetics
Cutoff %Subjects
Over Cutoff
Mean Liver Fat of
Cohort over
Cutoff
%Subjects Over Cutoff in
BMI>28 cohort (and Mean Liver Fat)
%Subjects Over Cutoff in
BMI>30 cohort (and Mean Liver Fat)
8% 11.5 13.9% 25.8 (14.2%) 30.6 (14.4%)
8% (in T2D) 45.2 14.9% 57.1 (15.5%) 60.2 (16.2%)
BMI, body mass index; T2D, type 2 diabetes mellitus.
Can we use BMI to screen for patients with high liver fat?
58. Understanding Medication and Liver Fat : Ex. Type II Diabetes
Controls 18%
Metformin Only 15%
Metformin + Pioglitazone 5%
Metformin + Gliclazide 8%
Metformin + Statins 46%
Gliclazide + Statins 8%
T2DM Patients with <5% Liver Fat (Normal): Visit 3
Controls 7%
Metformin Only 13%
Metformin + Pioglitazone 4%
Metformin + Gliclazide 18%
Metformin + Statins 45%
Gliclazide + Statins 13%
T2DM Patients with >5% (High) Liver Fat: Visit 3
Subjects with Liver Fat > 5%Subjects with Liver Fat < 5%
Sulfonylureas previously thought to
have a neutral effect on liver fat.
59. Next StepsâŚ
⢠Analysis using full imaging cohort
⢠Include additional parameters
available later this year
â Serum and urine biomarkers
â Health records with ICD10 codes
â Additional imaging biomarkers
⢠Liver MRI cT1 â indicator of fibrosis
⢠Carotid Ultrasound â
atherosclerosis indicator
⢠Increase focus to include âŚ
â Cardiovascular disease
â Muscle diseases
Blood data/samples
Urine data/samples
Genetic data
Questionnaire
Existing diseases
Health outcome
Death register
60. Take Home Points âŚ
⢠Radiomics provides insightful phenotyping.
⢠Imaging data, combined with other patient
data, can be mined with sophisticated
bioinformatics tools to develop models that
may potentially improve
â diagnostic,
â prognostic, and
â predictive accuracy.
⢠Radiomics could benefit numerous
therapeutic areas
61. Acknowledgements
Multidisciplinary data-mining efforts involve statisticians,
bio-informatists, geneticists, and other researchers.
Many Thanks to âŚ
⢠Melissa Miller - Genetics
⢠Joan Sopczynski â Predictive Informatics
⢠Yili Chen - Predictive Informatics
⢠Alexandra Dumitriu â Computational Biomedicine
⢠Craig Hyde â BioStatistics
⢠Jillian Yong â Imaging (Boston University)
And our Collaborators âŚ
⢠Jennifer Linge â AMRA Biostatistical Analyst
⢠Jimmy Bell â University of Westminster
63. Participants by socio-demographic factors
Characteristic Category Numbers (%)
Age 40-49 119,000 (24%)
50-59 168,000 (34%)
60-69 213,000 (42%)
Sex Male 228,000 (46%)
Female 270,000 (54%)
Ethnicity White 473,000 (95%)
Other 27,000 (5%)
Deprivation More 92,000 (18%)
Average 166,000 (33%)
Less 241,000 (46%)
Total 500,000
Generalisability (not representativeness): Heterogeneity of study
population allows associations with disease to be studied reliably
64. Build better software for life sciences
using user experience
The next Pistoia Alliance Discussion Webinar:
Moderator: Paula deMatos
Panel: Ewan Birney - Director of the EBI
Joel Miller - UX lead Amgen
Reed Fehr - Program Director, Customer Experience at idean
Date: December 5th, 2017 8am PT/11am ET/4pm GMT