Innovations in Digital Health: Exploring Advanced Biomarkers and AI Technologies

At AstraZeneca we harness data and technology to maximise time for the discovery and delivery of potential new medicines. Data science and artificial intelligence (AI) are embedded across our R&D to enable our scientists to push the boundaries of science to deliver life-changing medicines.

In this seminar, we will explore the cutting-edge advancements in digital health, focusing on the integration of AI and novel biomarkers to enhance diagnostic and therapeutic processes. The talk will cover a range of topics in the domain of machine perception.

The include vocal biomarkers, providing a non-invasive and accessible diagnostic tool for respiratory conditions. We will also discuss the latest developments in retinal imaging technologies and their applications in early detection of diseases such as diabetic retinopathy and age-related macular degeneration. Additionally, we’ll see advances in endoscopic techniques and AI-driven image analysis to improve the accuracy and efficiency of gastrointestinal diagnostics. We’ll also introduce EchoQC, a quality control system for echocardiography that leverages AI to ensure precise and reliable cardiac assessments.

Finally, we will present Evinova. AstraZeneca launched Evinova, set to be a leading provider of digital health solutions to better meet the needs of healthcare professionals, regulators and patients. With long-term backing from AstraZeneca and strategic collaborations with Parexel and Fortrea, Evinova offers globally-scaled digital products and services to the life sciences and healthcare sector.

1. Tuesday 4
th
March 2025
Innovations in Digital
Health: Exploring
Advanced Biomarkers
and AI Technologies
Tom Diethe
Head of the Centre for AI, DS&AI, Biopharma R&D, AstraZeneca

2. At AstraZeneca
we push the boundaries
of science to deliver
life changing medicines
Our purpose
is clear
3

3. 4
2bn
people with
BioPharma chronic
diseases
Top 5
causes of death
globally will be
CV disease, COPD,
CKD, and diabetes
by 2040
$22T
estimated
economic burden
from chronic
diseases and
COVID
(2011-2030)
1 in 3
adults have more
than one chronic
disease
* WHO COVID-19 Tracker

4. 5

5. 6
Addressing the biggest challenges in healthcare requires a shift from
symptom management to slowing and stopping disease
Precision medicine
approach
Target underlying
biology
New drug modalities
and combinations
Earlier detection,
diagnosis and
intervention
Time
Disease burden/progression
Severe
Death
Moderate
Mild
Healthy
Stop/
Reverse
Untreated
Slow
Symptomatic

6. Our transformation agenda will change
the trajectory of chronic diseases
AstraZeneca BioPharmaceuticals
7
Digital,
Data Science
& AI
Patient Centricity
Data Analytics,
Omnichannel and Go to
Market Models
End to
End Evidence
Innovative
Value Strategies
Better Predicting
Clinical Success
Next Generation
Therapeutics
Disease
Understanding
Pioneering
New Approaches
in the Clinic

7. Digital health
technologies can
reduce costs,
while improving
patient
experience and
outcomes
•Starting from around 6% of GDP in 2006-10, the combined public health and long-term
care expenditure for OECD countries is projected to reach at least 9.5% in 2060.
•In BRIICS countries, spending ratios will also increase significantly, reaching around 10%
of GDP by 2060 unless cost-containment policies are implemented.
•McKinsey estimates that digital health interventions alone have the potential to save
the US healthcare system nearly $500bn if fully adopted.
•Digital health is a large (+$900bn by 2032) and growing (13.6% CAGR 2022-2032)
market. The market for R&D digital health and care delivery with remote patient
monitoring, together make up approximately 60% of the total digital health market.
The remaining 40% consists of screening and diagnostics, wellness and disease
prevention, supply of therapies and digital pharmacies.
•A significant part of the cost and time of the drug development process is in the clinical
trials, rather than designing molecules. However, they are vital to the drug
development and approval process and critical in enabling access to innovative
medicines.
•Studies estimate that it now costs up to $2bn to bring a new drug to market and
according to one study the average length of time from the start of clinical testing to
approval is 7.1 years.
•Across the industry, almost 80% of clinical trials fail to meet recruiting timelines and
only 3-5% of eligible cancer patients join a clinical trial. Many groups, including
marginalised racial and ethnic groups, women, and other populations are
underrepresented in clinical trials.
•A recent AstraZeneca article published in Nature Medicine, lead author Cristina Duran,
demonstrated that the implementation of digital health technology in clinical trials can
improve the patient experience with accelerated timelines and reduced costs.
8
Durán, Cristina Ortega, et al.
"Implementation of digital health
technology in clinical trials: the 6R
framework."nature medicine29.11
(2023): 2693-2697.

8. 9
Opportunity Risks Regulation
We are innovating in an emerging field with great
opportunity, inherent risks, and evolving regulation

9. AI and Machine Learning are now embedded in science
10

10. We follow five principles for ethical AI
11
Private and SecureExplainable and
Transparent
Fair Human-CentricSocially Beneficial
AstraZeneca bioethics policy:
https://www.astrazeneca.com/sustainability/resources.html#policies-and-standards-0

11. 12
FAIR data is at the heart of our AI ambitions
Analytics-ready data drives scalable, and
efficient data science operations enabling
efficient & seamless experience for
scientists to access and analyse their
data.
We’re applying a targeted effort to the
curation and development of FAIR-tools,
to make data more FAIR and higher value,
for use and re-use.

12. 13
Flavours of AI
AI for ProductivityAI for Science
Generative AI – "Large
Language Models"
Document/image
summarisation/generation
Information Extraction
AI for Software Engineering
Virtual Assistants
Knowledge Gathering for Business Development
…
Target discovery and validation
Molecular property prediction and generation
Identification of disease-causing genomic variants
Improve CRIPSR efficiency
Optimal experimental design
…

13. Our ambitions are
made possible by our people
and our skills

14. Centre for Artificial Intelligence: A Global Team
15
Gaithersburg
Cambridge
Gothenburg
Barcelona
We are a global, science-led, patient-focused pharmaceutical company.
We are dedicated to transforming the future of healthcare by unlocking
the power of what science can do for people, society and the planet.
Mississauga

15. X-Industry Biotech Fund
✓Creating and accelerating
start-ups for R&D with AI
with 7 startups founded to
date.
✓Aiding antibody discovery &
optimization, clinical
outcome prediction
16
Academic Collaboration
✓Tapping into
interdisciplinary,cutting-
edge innovation
✓Building up from 10 to 20-25
PhD studentships
Industry Collaboration
✓Integrated AI and wet lab
technologies to design
developable, selective, and
functionally active protein
drugs against multipass
membrane protein targets
We use coordinated collaborations to create value across R&D

16. Case Studies
17

17. Agenda
18
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

18. Agenda
19
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

19. Cough Detection with Vitalograph
•Vitalograph specialises in solutions
for cough detection
•Cough detection aids in diagnosing
respiratory conditions
•The Vitalojak (pictured) is the only
FDA approved medical device for
cough detection
20

20. Challenge:
Increase the cost
efficiency of using
cough as an endpoint in
trials for indications
with underlying cough,
such as COPD, and
eventually to support a
chronic cough
indications
21
Cough is a recurring symptom and represents
a frequent patient complaint in many of the
respiratory diseases in which we are
developing therapies.
Cough in trials can be monitored through
audio recording, and the number of coughs in
24 hours can be assessed manually, i.e. a
human is paid to listen and manually update a
count of coughs
Seek to use AI to reduce the amount of a
recording that has to be manually reviewed.

21. Accurate recall for cough detection across all frequency ranges, with
precision increasing with cough frequency (coughs [cs] per 24 hour [h])
•Developed and retrospectively validated AI cough
detection by comparing with cough labelling by
humans from COPD and Asthma trial data
•We can reduce manual audio review time by half (to an
average of just 33 minutes) with negligible loss in
cough events, outperforming the leading commercially
available option.
•This method has an average recall of 97% and precision
of 57% for cough detection using more than 500 24-
hour audio recordings.
•Additionally, we have shown that our method can be
more consistent than human counters by assessing
inter-rater reliability of human counts, revealing a 95th
quantile human measurement error of 7.4%.
22

22. Minimal Viable Product
•Successfully developed a Cough
Detection Web Application
•Instant Analysis: Upload audio files
and our AI model starts analysing.
•Reporting Results: Returnsaudio
file and highlights regions with
cough occurrences.
23

23. Agenda
24
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

24. Retinal Imaging for Cardiovascular Health
•Retinal imaging provides non-
invasive insights into cardiovascular
conditions.
•It helps in early detection of
diseases like diabetes and chronic
kidney disease.
•Linking retinal health to overall
cardiovascular endpoints enhances
patient care.
25

25. Its time to change the way we do things; retinal imaging pre-screening has
potential to save time, cost & carbon efficiencies
Cohort enrollment
Retinal pre-screening
Earlier detection
of screen failures
wrt BP, eGFR
Pre-screening tool development for internal use
•Collaboration at Ph1 study units in mainstream programs e.g. early CVRM studies to evaluate model performance in screen
failure detection wrt BP and eGFR. Safety follow up assessment for screen volunteer incorporated
•Pilot initially and then proceed to prospective evaluation to support use in future Ph1 and other studies
•Approval as diagnostic will not be required as readings will be repeated with regulatory accepted measures
•Camera cost ~ 10k $ incl cloud storage/site, API
Screening
Screen failures in Ph1 ≈
10% due to BP & eGFR,
at a cost of 3k$/patient
Study sites
CVRM
Ph1 studies
Study visit

26. 27
Healthy Heart Africa : increase operational efficiency, build sustainability
and access to health care. Validation in Egypt, Uganda, Kenya
Pre-screening
facilitation
Companion
Diagnostic
Diagnosis &
Prescribing
algorithm
Increase operational efficiency
•Decreasing screen failure rate
•Potential time and cost savings
•Enhance patient identification
Build sustainability and access to health
care
•Simplified and remote diagnosis and
prescriptions
•Decrease CO2 footprint via
consumables, travel

27. Machine-Learning Based Prescription: Our Vision
Machine-learning
predictions
HbA1c/T2DM
diagnosis
SBP/DBP
Retinal Image
CKD & T2DM
Management
eGFR
Diagnosis of CKD
•Confirm diagnosis
• Assess risk
• Decide management
CV risk

28. Overview of end-end process for retinal image acquisition
and machine-learning based analysis
ECP, eye care professional; eGFR, estimated glomerular filtration rate
Flag positive risk
No action
Matrix-Entropy
QC filter
Age
Gender
Macular centered
retinal fundus
photographs
BYOL pretrained EfficientNetB1
Convolutional Neural Network
NGBoost Binary Classifier
1
OR
1. Duan, T. , Anand, A., Ding, D.Y ., Thai, K .K., Basu, S ., Ng, A. and S chuler, A., 2020, November. Ngboost: Natural gradient boost ing for
probabilistic prediction. InInternational Conf erence on Machine Learning(pp. 2690-2700). PMLR.
Output
Features considered
Preprocess Images
[RoI Crop, CLAHE, Norm]
Laterality
Entropy
Visual
feature
vector

29. Refinement and improvement of predictive performance
for CKD ( eGFR above or below 60 ml/min/1.73m2)
AUC, area under the curve; CKD, chronic kidney disease; FP, false positive; QC, quality control; ROC, receiver operating characteristic
Keshvari-Shad, Fatemeh, et al. "A systematic review of screening tests for chronic kidney disease: an accuracy analysis."Galen Medical Journal9 (2020): e1573.
‘21 May ‘22July ‘22June ‘24
Current stage 3, binary CKD ROC AUC
Prediction performance (Dev Testing)
0.630.650.74~0.85
CKD prediction
QC filtering
‘Scale-targeted entropic
filtering’ automate QC
Filtering offers yield versus
performance calibration
(reducing burden of FPs)
Packaging
Dockerized and on track for
packaging and deployment
*UACR prediction performance varies depending on the quality of the testing strip; For the detection of ACR>30 mg/g, the sensitivities of the dipstick with a cut-
off point of trace were ranged from 37.1-69.4% and specificities from 93.7-97.3%

Straight Calibration curve ROC performance Interpretability plots

30. Agenda
31
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

31. Endoscopy is a minimally invasive procedure for internal
examination
•It uses a flexible tube with a light and
camera to visualize organs.
•Endoscopy aids in diagnosing and
treating various medical conditions.
•Common types include gastrointestinal,
bronchial, and urological endoscopy.
•The procedure allows for real-time
imaging and intervention.
32
This Photo by Unknown Author is licensed under CC BY-SA

32. Challenge:
Human expert
disagreement, capture
quality variability, and lack
of a clear decision
boundary while manually
grading Ulcerative Colitis
(UC), based on endoscopy
exploration
Loss of precision in a key-
endpoint parameter is a
concern for related
studies.
33
Ulcerative Colitis isan
inflammatory bowel disease (IBD)
that causes inflammation and
ulcers in the digestive tract.
Human experts grade UC using the
Mayo score, which includes an
endoscopy subcomponent with
scores from 0 to 3.
Intended uses are clinical proof-of-
concept and registration of new
drugs.

33. 34
Self-Supervised Multi-instance learning to score endoscopy
videos using video-level annotations

34. 35
EndoAI Results
Mean AUC-
ROC
Std AUC-ROC
Severe model
0.83 0.03
Healing model
0.85 0.02
Healthy model
0.81 0.01

35. Agenda
36
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

36. Echocardiograms
•Echocardiograms use sound waves
to create heart images.
•They assess heart function and
structure non-invasively.
•Echocardiograms help diagnose
various heart conditions.
•Left Ventricle Ejection Fraction
(LVEF) is a key biomarker
37

37. Challenge:
Apical view foreshortening
(FS) is a commonly
occurring
and hard to detect source
of variable quality in echo
capture in clinical trials.
FS arises from variations
in placement of the
ultrasound probe
during capture and can
result in over-/under-
estimation of critical
endpoints such as
LV volumes, ejection
fraction or strain
38
A real-time system capable of prompting
sonographers, at the point of capture, to adjust
the probe position and achieve the highest
quality recording, would have significant
positive impact on data quality in future trials

38. Real-time cardiac predictions – Inception-based model
•Convolution blocks (gray boxes) are composed of
convolutions, batch normalization and PReLU
activations.
•Two versions of the Inception module are employed:
the illustrated one being used in the lower part of the
network (dark purple) and a simplified one without the
5×5 route in higher parts of the network (bright
purple).
•The final classifier block consists of another
compressing convolution layer with kernel size 1×1
and filter amount equal to the number of views. The
output is activated with a PReLU layer. Finally, global
average pooling followed by softmax activation yields a
prediction vector as output.
•Runtime >40 FPS on Apple M2 silicon
39
Smistad et al 2020. Real-time automatic ejection fraction and
foreshortening detection using deeplearning.IEEE transactions on
ultrasonics, ferroelectrics, and frequency control,67(12), pp.2595-2604.

39. 40
18.1% Reduction in
Mean Absolute Error
For left ventricular
ejection fraction (LVEF)
EchoQC – reducing error in endpoints and screening.
Next, we will assess the impact our QC could have on other
clinical parameters derived from echocardiography.

40. Agenda
41
Cough Detection
Retinal Imaging for CV Health
Endoscopy Videos
Echocardiograms
Evinova
1
2
3
4
5

41. Evinova
•In November 2023, AstraZeneca launched Evinova, a
health-tech business to accelerate innovation across
the life sciences sector, the delivery of clinical trials and
better health outcomes
•Evinova operates as a separate health-tech business
within AstraZeneca
•Evinova’s globally-scaled digital health solutions are
evidence-led, science-based and human experience-
driven to serve clinical trial sponsors, clinical
research organisations (CROs), clinical trial site care
teams and patients
•First strategic collaborations with globally-leading
CROs Parexel and Fortrea enable Evinova’s digital
health solutions to be offered to their wide
customer base
42

42. Conclusions
•Digital health innovations are
transforming patient care.
•AI technologies enhance
diagnostics and treatment
accuracy.
•Continued collaboration is essential
for future advancements.
43

43. Questions?
44

44. Confidentiality Notice
This file is private and may contain confidential and proprietary information. If you have received this file in error, please
notify us and remove it from your system and note that you must not copy, distribute or take any action in reliance on it.
Any unauthorized use or disclosure of the contents of this file is not permitted and may be unlawful.
AstraZeneca PLC, 1 Francis Crick Avenue, Cambridge Biomedical Campus, Cambridge, CB2 0AA, UK
[phone]
www.astrazeneca.com
45