Les technologies digitales et l’automation ont permis d’améliorer la productivité dans de nombreux domaines industriels, mais peu encore en médecine. Le concept de digital twins s’est construit à l’intersection de l’industrie 4.0 et de l’internet des objets, et désigne une copie virtuelle d’un système capable d’interagir de manière bidirectionnelle et en temps réel avec le système physique. En médecine, le digital twin est souvent utilisé comme un synonyme de modèles numériques personnalisés et n’ayant pas toujours le couplage en temps-réel avec le système physique. Le digital twin utilise les connaissances venant de la recherche biomédicale et peut se nourrir également des données cliniques. Il permet de donner une réponse à des questions thérapeutiques, allant de la découverte de nouvelles thérapies jusqu’à la prédiction de l’effet d’une intervention thérapeutique proposée. Après une introduction du concept de digital twins, nous illustrerons son application à la médecine personnalisée à l’aide d’exemples en médecine régénérative et aux soins intensifs.

1. Vendredi, 26 novembre 2021
Le concept de digital twins appliqué à la médecine personnalisée
Liesbet Geris & Thomas Desaive (GIGA in Silico Medicine, ULiège)

2. LIEGE CREATIVE, en partenariat avec :

3. Le concept de digital twins appliqué à la
médecine personnalisée
From bench to ICU bedside
Thomas Desaive
GIGA – In silico medicine
LIEGE CREATIVE, 26/11/21

4. Context

5. ▶ Healthcare costs are 10-17%
of GDP in western/modern
healthcare countries
▶ Growing at a rate of 7-8% a
year (doubles every 10 years)
Economics
Health spending (%GDP)
(source: OECD data)
Health spending (USD/capita)
(source: OECD data)

6. ▶ Healthcare costs grow faster than GDP
▶ Ability to deliver equity of access to care under stress
▶ Aging demographics and chronic disease multiply the effects
▶ Increasing need to reduce the cost of care so we can cover costs of care rising faster than the ability
to pay
à Improve productivity of care
Healthcare
Digital technologies and automation have brought significant productivity gains to many industries, and
manufacturing in particular.
However, such productivity gains have not yet come to the field of medicine...

7. In silico medicine

8. In vitro, in vivo… in silico!
Use of computer simulation in the understanding, diagnosis, treatment, or prevention of a disease
à modeling, simulation, and visualization of biological and medical processes in computers with the
goal of simulating real biological processes in a virtual environment.
In silico medicine

9. Digital twins

10. “A virtual copy of a system able to interact with the physical system in a bi-directional way”
What’s a digital twin?
▶ Key definitions:
- Digital Model (DM) = Model with no
interaction with the physical system
- Digital Shadow (DS) = Model receives data
from physical system and updates
- Digital Twin (DT) = Model receives data,
updates, and sends data or control back to the
physical system (patient)
▶ Bi-directional exchange of information synchronizes
virtual system response to match the physical
system to “forecast and optimise the behavior of
the physical system in real time”.

11. From manufacturing to medicine…
▶ On top of a “control layer” of supporting
technologies (IoT, sensors, cloud computing…)
▶ Under a top layer of ”entreprise resource planning”
guiding how the DT is applied
à Protocols and hospital care guidelines
▶ The upper and lower layers both inform the DT and
its design and use.
▶ The middle layer DT itself is connected through the
control layer to its physical counterpart (the patient)
in real-time and uses modeling and computation to
continually update the DT

12. Digital twins in personnalized medicine

13. Digital twin creation

14. Applications in
intensive care

15. ▶ Intensive care units (ICU) consume 1-2% of GDP!
▶ ICU patients are very complex, highly variable and difficult to
manage
▶ Number of ICU beds limited!
▶ Technology is everywhere (pumps, ventilators, monitors…)
▶ Lack of interoperability between devices
à Digital twins requirements
▶ Physiologically relevance
▶ Clinical relevance
▶ Treatment sensitivity, practically identifiable from clinically
available data (no new sensors added if possible).
Specificities

16. Glucose control in ICU

17. ▶ Critically ill patients often exhibit hyperglycemia due to the stress-response, even with no prior
history of diabetes
▶ Several studies have linked hyperglycaemia in critically ill patients with increased mortality and
worsened outcomes
- 17-45% reductions in mortality by controlling BG to ‘normal’ levels
- Reduced sepsis, organ failure, heart attacks and more
▶ BUT
- Few groups have managed to replicate the results of the initial ‘landmark’ studies by Van den
Berghe et al. and Krinsley et al
- Glycaemic control turns out to be fairly difficult
› Fear of hypoglycemia
› Humans are horribly variable! – Inter and intra-patient variability, changes over time and
with treatment etc...
Why caring about glucose control in ICU?

18. ▶ The main issues:
- Variability: Inability to control glucose to target tightly
- Variability: Increased hypoglycemia
- Variability: Tendency to measure infrequently
▶ Solution?
- Model-based methods that identify patient condition and adapt treatment to match
- Model-based methods that understand the likely variability in patient condition and
response to therapy
- Model-based engineering of clinical therapy, as developed from engineering models
and methods
The problem and solution (?) Poor performance
and inablity
to achieve outcomes
Digital twin!

19. Variability, not physiology or medicine…
Models offer the opportunity to identify, diagnose and
manage variability directly, to guaranteed risk levels.
Blood
Glucose
levels
Controller
Fixed dosing systems
Typical care
Adaptive control
Engineering approach
Variability flows
through to BG control
Variability stopped at
controller
Fixed protocol treats
everyone much the
same
Controller identifies and
manages patient-specific
variability
Patient
response to
insulin

20. ▶ Insulin sensitivity is a key feature of these models
- All the models that have been successfully used for glycemic control include one or
more insulin sensitivity parameters
- The insulin sensitivity parameter describes/captures the insulin mediated transport of
glucose across the cell membrane
- This form of model mimics the physiological behavior
The physiology?
Free insulin binds to receptors on the cell
wall, triggering a signalling cascade that
results in GLUT4 transporters moving to the
cell wall. GLUT4 actively transports glucose
into the cell

21. ▶ Physiological glucose-insulin system model (Chase et al.)
Modeling
3
2
)
(
1
1
max
2
1
max
2
)
(
1
.
)
,
2
.
min(
)
(
2
)
(
1
.
)
(
1
)
(
)
,
2
.
min(
)
(
)
(
)
1
(
)
(
))
(
)
(
(
)
(
1
)
(
)
(
.
)
(
)
(
1
)
(
))
(
)
(
(
)
(
)
(
)
(
1
)
(
)
(
).
(
)
(
.
)
(
k
t
I
en
I
en
L
I
ex
I
I
L
K
G
C
I
G
G
G
k
e
k
I
u
P
d
P
P
d
t
P
t
D
P
d
t
P
t
PN
P
P
d
t
P
V
I
u
x
V
t
u
t
Q
t
I
n
t
I
t
I
n
t
I
n
t
I
t
Q
t
Q
n
t
Q
t
I
n
t
Q
V
CNS
EGP
t
P
t
Q
t
Q
t
G
t
SI
t
G
p
t
G
=
+
-
=
+
-
=
+
=
-
+
+
-
-
+
-
=
+
-
-
=
-
+
+
+
-
-
=
!
!
!
!
!
a
a
a
The insulin sensitivity parameter defines
insulin-mediated glucose uptake
A compartment model composed
of several coupled sub-models

22. Determining insulin sensitivity
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Patient data from the hospital records:
BG measurements
Dextrose Insulin infusion

23. Determining insulin sensitivity
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Patient data from the hospital records:

24. Determining insulin sensitivity
Patient data from the hospital records:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]

25. Determining insulin sensitivity
Patient data from the hospital records:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Model BG fit
Identified insulin sensitivity

26. Virtual trials
0 2000 4000 6000 8000 10000 12000
0
5
10
15
Blood
Glucose
Retrospective
Simulated control
0 2000 4000 6000 8000 10000 12000
0
0.2
0.4
Insulin
Input
[U/kg/hr]
0 2000 4000 6000 8000 10000 12000
0
50
100
I,
Q
[mU/L]
0 2000 4000 6000 8000 10000 12000
0
20
40
60
Feed
mg/kg/min
0 2000 4000 6000 8000 10000 12000
0
0.5
1
1.5
x 10
-3
sI
0 2000 4000 6000 8000 10000 12000
0
5
10
15
Blood
Glucose
Retrospective
Simulated control
0 2000 4000 6000 8000 10000 12000
0
0.2
0.4
Insulin
Input
[U/kg/hr]
0 2000 4000 6000 8000 10000 12000
0
50
100
150
I,
Q
[mU/L]
0 2000 4000 6000 8000 10000 12000
0
20
40
60
Feed
mg/kg/min
0 2000 4000 6000 8000 10000 12000
0
5
10
15
Retrospective
Simulated control
0 2000 4000 6000 8000 10000 12000
0
0.5
1
0 2000 4000 6000 8000 10000 12000
0
100
200
300
0 2000 4000 6000 8000 10000 12000
0
5
10
15
Blood
Glucose
0 2000 4000 6000 8000 10000 12000
0
0.2
0.4
Insulin
Input
[U/kg/hr]
0 2000 4000 6000 8000 10000 12000
0
50
100
I,
Q
[mU/L]
0 2000 4000 6000 8000 10000 12000
0
20
40
60
Feed
mg/kg/min
0 2000 4000 6000 8000 10000 12000
0
0.5
1
1.5
x 10
-3
sI
Retrospective
Actual BG control
Actual insulin/nutrition given
Virtual patient
from hospital
records
Simulation results
… and more …
Gave more insulin at a certain time?
Gave less insulin?
Measured BG at a different time?
Test ideas on this virtual
patient on computer
What would have happened if….
Compare to actual BG control above
Actual
patient

27. Digital twin for glucose control in ICU
Standard infuser
equipment adjusted by
nurses
Patient management
Measured data
Decision Support
System
I
en
L
I
ex
I
K
I
L
G
c
I
G
b
G
I
G
V
G
u
x
V
t
u
t
Q
t
I
n
t
I
n
t
I
t
I
n
I
t
Q
t
Q
n
t
Q
t
I
n
Q
V
t
PN
CNS
EGP
P
P
d
t
Q
t
Q
t
G
S
t
G
p
G
)
(
)
1
(
)
(
))
(
)
(
(
)
(
)
(
1
)
(
)
(
1
)
(
))
(
)
(
(
)
(
)
,
min(
)
(
1
)
(
)
(
)
(
.
.
max
2
2
.
-
+
+
-
-
-
+
-
=
+
-
-
=
+
-
+
+
+
-
-
=
a
a
a
Identify and utilise
“immeasurable”
patient parameters
For insulin
sensitivity (SI)

28. ▶ The insulin sensitivity metric tells us how a patient
has been responding to insulin recently.
▶ BUT...we need some idea of how the patient will
respond into the future...
- A sudden increase in insulin sensitivity may
result in hypoglycaemia...
▶ Stochastic modelling using retrospective data
indicates the likelihood and magnitude of changes
in insulin sensitivity
- Quantify the likely range of insulin sensitivity
for the upcoming 1 & 2 hours
- A bit like forecasting the weather…
Real-time clinical control
2D Kernel density model of
insulin sensitivity
Improved control performance and reduced risk

29. Stochastic model in action
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Stochastic model:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]

30. Stochastic model in action
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Stochastic model:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Insulin sensitivity might not
change much, so expect a
~constant BG response

31. Stochastic model in action
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Stochastic model:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Insulin sensitivity might rise
suddenly, so there is a
possibility of lower BG

32. Stochastic model in action
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Stochastic model:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Insulin sensitivity might
drop suddenly, so there
may spike in BG

33. Stochastic model in action
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
1
2
x 10
-3
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Stochastic model:
0 20 40 60 80 100
0
5
10
15
Patient 7
BG
[mmol/L]
0 20 40 60 80 100
0
0.5
1
S
I
[L/(mU.min)]
0 20 40 60 80 100
0
0.02
0.04
0.06
Time [hours]
Dextrose
[mmol/min]
0 20 40 60 80 100
0
0.5
1
1.5
Insulin
[mU/min]
Work out the 90%
confidence range for future
insulin sensitivity and BG
values.
Match the forecasted range
with our blood glucose
targets.

34. Example
Green area indicates 90%
probability forecast
Goal of BG targeting is to
ensure bottom edge of green
line is above 4 mmol/L
à Thus, risk of hypoglycaemia
is incorporated directly into
BG control.

35. Clinical results

36. ▶ Bi-directional exchange of data between the digital twin and the patient is the key
- Need access data to medical devices
- Need greater interoperability between medical devices
▶ Adoption of a digital twin is at the « entreprise layer » and requires a decision at the ICU level, the
hospital and/or health system level
- Patients are reported to benefit from only 30-50% of the validated healthcare technologies
- Adoption requires consideration of factors at the individual, the unit and the system levels
(technology, ergonomics, ability to use the innovation)
▶ Need for social sciences to enhance adoption at the individual level but also at the organisation level
(hierarchical nature of healthcare)
Why not more digital twins in medicine?

37. Thanks to…

38. Questions?

39. Digital Twins for
Personalised Heathcare
Liesbet Geris

40. Digital Twins in personalized healthcare
▶ Digital Twin
- Concept from Industry 4.0
- DT = virtual representation of a
physical object or system across its
life-cycle. It uses real-time data
and other sources to enable
learning, reasoning, and
dynamically recalibrating for
monitoring, diagnostics and
prognostics.
2
Copyrighted images Copyrighted images

41. All phases of the R&D pipeline
www.fda.gov (CDRH report) 3

42. And beyond
www.fda.gov (CDRH report) 4

43. Digital twins for knee OA treatment design
Disease-modifying
drugs
30/11/2021 5
Tissue Engineering
strategies
Total Knee Replacement

44. Lesage et al;, in preparation, 2021
Identifyig drugable targets

45. Optimising 3D culture process
30/11/2021 All rights reserved © 2020 7
Top
Side
Bottom
Guyot et al., BMMB, 2016
Scaffold:
Ø = h= 6 mm

46. Mandal et al;, in preparation, 2021
Optimising printing process

47. Vanhede et al., Adv Funct Mat, 2021
Optimising scaffold design

48. 3D scaffolds implanted in cranial augmentation model (rat, n=10 per condition)
Optimising scaffold design
30/11/2021 10
5 mm
Vanhede et al., Adv Funct Mat, 2021

49. ▶ In silico approaches contribute to
- Increase understanding of pathophysiology
- Design treatment strategies
▶ In silico approaches part of R&D pipeline
- Quality control
- 3R’s
- Personalisation
▶ QUESTION: how to build credibility?
Summary
30/11/2021 All rights reserved © 2020 11

50. © 2021 VPH Institute
How to build credibility?
• Key elements
– Documentation
– VVUQ
– Standardization
• Regulatory approval
– Digital evidence
– Software as a medical
device / tool

51. © 2021 VPH Institute
13

52. © 2021 VPH Institute
VVUQ
• Verification
– Does the computational model accurately solve the
underlying mathematical model?
• Validation
– How well does the computational
model approximate ‘reality’?
• Uncertainty Quantification
– How much does uncertainty in
parameters/assumptions affect the simulation results?
14
Source: Tina Morrison, FDA
Mathematical
model
Reality
Numerical
Solution
Assumptions
Mathematical analysis
Numerics
Software
Parameters
Assumptions

53. © 2021 VPH Institute
Standardisation
• Computational Modeling of Medical Devices
– ASME V&V40 Subcommittee is a standards
organization with more than 60 industry partners
• New standard published in 2018:
– Assessing Credibility of Computational Modeling
through Verification and Validation: Application
to Medical Devices
15

54. © 2021 VPH Institute
Risk identification in V&V40
16
ASME - V&V40 FDA temporary regulatory framework of AI/ML-based SaMD

55. © 2021 VPH Institute
Regulatory framework for AI
• Pre-
determined
Change
Control Plan
• Algorithm
Change
Protocol

56. © 2021 VPH Institute
EU? Context & ongoing initiatives
• ASME V&V40
• GSP initiatives
• EMA-HMA joint network strategy to 2025
• ICH MIDD reflection

57. © 2021 VPH Institute
Mandate from EC
• EC’s Regulation Proposal to extend the mandate of the EMA :
– “In order to facilitate the work and the exchange of information under
this Regulation, provision should be made for the establishment and
management of IT infrastructures and synergies with other existing IT
systems or systems under development, including the EUDAMED IT
platform for medical devices. That work should also be facilitated by,
where appropriate, emerging digital technologies such as
computational models and simulations for clinical trials, as well as
data… ”

58. © 2021 VPH Institute
Concept of regulatory impact

59. © 2021 VPH Institute
White paper published

60. © 2021 VPH Institute
Triple Helix Expertise Exchange meeting
• Organised by VPHi
• Aim:
– Exchange expertise
– Between regulators, academia & industry
– On specific in silico topics

61. © 2021 VPH Institute
Good Simulation Practice
• Need for guidance for in silico technologies
– Good examples from FDA
• Growing need to agree on a Good Simulation Practice
document widely recognized and accepted
– Like the Good Clinical Practice
• Effort towards global harmonization
– But granularity is not clear
https://insilicoworld.slack.com

62. © 2021 VPH Institute
From the screen to the patient:
a community effort
Policies
Incentives
Community
User experience
Technical implementation

63. 25