Departm e nt of Inten sive C are

Geoffrey M Shaw1

J Geoffrey Chase2

Balazs Benyo3

1 Dept of Intensive Care, Christchurch Hospital2 Dept of Mechanical Engineering, Univ of Canterbury3 Dept informatics, Budapest University of Technology

and Economics

Model-based Therapeutics: Tomorrow’s care at yesterday’s cost

NZ ANZICS Dunedin March 15 2013

The bread and butter of ICU:

Some of the basic things that we do...

• Glucose control and nutrition• Sedation• Cardiovascular management: “tropes and fluids”• Mechanical ventilation

The bread and butter of ICU:Intuition and experience, provides the fundamental basis of care delivered to the

critically ill; it is specific to the clinician, but it is not specific to the patient.

The result: highly variable and over customised carepoor quality and increased costs of care,

What are needed :

Treatments that are patient specific and independent of clinician variability and bias

A “one model”, not “one size”, fits-all approach

The bread and butter of ICU:

• Glucose control and nutrition• Mechanical Ventilation (next presentation!)

Model based therapeutics “MBT”

Model based therapeutics “MBT”First, we describe the physical systems to

analyse

Model based therapeutics “MBT”Next, we build up a

mathematical representation of the system

I

enL

I

exIK

I

L

GcI

G

b

GIG

VGu

xVtu

tQtIntIntItIn

I

tQtQntQtInQ

VtPNCNSEGPPPd

tQtQtGStGpG

)()1(

)())()(()(

)(1)(

)(1)())()((

)(),min()(1

)()()(

.

.

max22.

Model based therapeutics “MBT”

I

enL

I

exIK

I

L

GcI

G

b

GIG

VGu

xVtu

tQtIntIntItIn

I

tQtQntQtInQ

VtPNCNSEGPPPd

tQtQtGStGpG

)()1(

)())()(()(

)(1)(

)(1)())()((

)(),min()(1

)()()(

.

.

max22.

Finally, we use computational analysis to solve these equations to help us design

and implement new, safer therapies.

I

enL

I

exIK

I

L

GcI

G

b

GIG

VGux

VtutQtIntIn

tItInI

tQtQ

ntQtInQ

VtPNCNSEGPPPd

tQtQtGStGpG

)()1()())()(()()(1

)()(1

)())()((

)(),min()(1

)()()(

.

.

max22.

So where does this go?

Doctors clinical experience

and intuition

Insulin Glucose Sedation Steroids and vaso-pressors Inotropes And many many more …

• Glucose levels• Cardiac output• Blood pressures• SPO2 / FiO2• HR and ECG• And many more…

Insulin Sensitivity Sepsis detection Circulation resistance

A better picture of the patient-specific physiology in real-time at the bedside

Optimise glucose control

Manage ventilation Diagnose and treat

CVS disease And many other

things…

http://upload.wikimedia.org/wikipedia/commons/8/8c/Insulin_glucose_metabolism.jpg

A wish list• What will happen if I add more insulin?

• What is the hypoglycemia risk for this insulin dose?– Over time?– When should I measure next to be sure?

• How good is my control? Does it need to be better?

• Should I change nutrition? What happens if someone else has changed it? How should I then change my insulin dose?– Many if not all protocols are “carbohydrate blind” and thus BG is a very poor surrogate of

response to insulin

• Is patient condition changing? What happens if it changes between measurements?

Standard infuser equipment adjusted by nursesPatient management

Measured data

“Nurse-in-the-loop” system. Standard ICU equipment and/or low-cost commodity hardware.

Decision Support System

I

enL

I

exIK

I

L

GcI

G

b

GIG

VGux

VtutQtIntIn

tItInI

tQtQ

ntQtInQ

VtPNCNSEGPPPd

tQtQ

tGStGpG

)()1()())()(()()(1

)()(1

)())()((

)(),min()(1

)()()(

.

.

max22.

Identify and utilise “immeasurable”

patient parametersFor insulin sensitivity

(SI)

Feedback control

ICU bed setup

Nutrition pumps:Feed patient through nasogastric tube, IV routes or meals

Glucometers:Measure blood sugar levels

Infusion pumps:Deliver insulin and other medications to IV lines. Sub-cut insulins may also be used.

INPUT OUTPUT OUTPUT

Blood Glucose levels

Controller

Fixed dosing systemsTypical care

Adaptive controlEngineering approach

Variability flows through to BG control

Variability stopped at controller

Models offer the opportunity to identify, diagnose and manage variability directly, to guaranteed risk levels.

Fixed protocol treats everyone much the same

Controller identifies and manages patient-specific

variability

Patient response to

insulin

Variability, not physiology or medicine…

BG [mg/dL]

Time

4.4

6.5

Insulin sensitivity

Blood glucose

tnow

Stochastic model shows the bounds (5th – 95th percentile) for insulin sensitivity variation over next 1-3 hours from the initially identified level

For a given feed+insulinintevention an output BG distribution can be forecast using the model

tnow+(1-3)hr

95th

75th

50th

25th

5th

5th

25th

50th

75th

95th

5th, 25th, 50th (median), 75th, 95th percentile bounds for SI(t) variation based on current value

Insulin sensitivity

Blood glucose

tnow

Stochastic model shows the bounds (5th – 95th percentile) for insulin sensitivity variation over next 1-3 hours from the initially identified level

For a given feed+insulinintevention an output BG distribution can be forecast using the model

tnow+(1-3)hr

95th

75th

50th

25th

5th

5th

25th

50th

75th

95th

Stochastic model predicts SI

Forecast BG percentile bounds:

A predicted patient response!

SI percentile bounds

+known insulin

+system model

= ...

Iterative process targets this BG forecast to the range we want:

= optimal treatment found!Patient response forecast can be recalculated for

different treatments

Models, Variability and Risk

BG [mg/dL]

Time

4.4

6.5

Insulin sensitivity

Blood glucose

tnow

Stochastic model shows the bounds (5th – 95th percentile) for insulin sensitivity variation over next 1-3 hours from the initially identified level

For a given feed+insulinintevention an output BG distribution can be forecast using the model

tnow+(1-3)hr

95th

75th

50th

25th

5th

5th

25th

50th

75th

95th

5th, 25th, 50th (median), 75th, 95th percentile bounds for SI(t) variation based on current value

Insulin sensitivity

Blood glucose

tnow

Stochastic model shows the bounds (5th – 95th percentile) for insulin sensitivity variation over next 1-3 hours from the initially identified level

For a given feed+insulinintevention an output BG distribution can be forecast using the model

tnow+(1-3)hr

95th

75th

50th

25th

5th

5th

25th

50th

75th

95th

Stochastic model predicts SI

Forecast BG percentile bounds:

A predicted patient response!

SI percentile bounds

+known insulin

+system model

= ...

Iterative process targets this BG forecast to the range we want:

= optimal treatment found!Patient response forecast can be recalculated for different treatments

Maximum 5% Risk of BG < 4.4 mmol/L

Why this approach?• Model lets us guarantee and fix risk of hypo- and hyper- glycemia

• Giving insulin (and nutrition) is a lot easier if you know the range of what is likely to happen.

• Thus, one can optimise the dose under all the normal uncertainties– No risk of “unexplained” hypoglycemia

• Allows clinicians to select a target band of desired BG and guarantee risk of BG above or below

• We tend to fix a 5% risk of BG < 4.4 mmol/L which translates to less than 1/10,000 (interventions) risk of BG < 2.2 mmol/L (should be about 2% by patient)– Fyi, this is how airplanes are designed and how Christchurch's high rises should

have been designed!

Some Results to Date• Very tight

• Very safe

• Works over several countries and clinical practice styles

• Also been used in Belgium

• Measuring SI is very handy whether you do it with a model (STAR) or estimated by response (SPRINT)

STAR Chch STAR Gyula SPRINT Chch SPRINT Gyula

Workload# BG measurements: 1,486 622 26,646 1088Measures/day: 13.5 12.8 16.1 16.4Control performance

BG median [IQR] (mmol/L): 6.1[5.7 – 6.8]

6.0[5.4 – 6.8]

5.6[5.0 – 6.4]

6.30[5.5 – 7.5]

% BG in target range)* 89.4 84.1 86.0 76.4% BG > 10 mmol/L 2.48 7.7 2.0 2.8Safety% BG < 4.0 mmol/L 1.54 4.5 2.89 1.90% BG < 2.2 mmol/L 0.0 0.16 0.04 0

# patients < 2.2 mmol/L 0 1 (started hypo) 8 (4%) 0

Clinical interventionsMedian insulin (U/hr): 3 2.5 3.0 3.0

Median glucose (g/hr): 4.9 4.4 4.1 7.4

*4-8mmol/L

So, because we know the risk …

• We get tight control

• We are very safe

• We do it by identifying insulin sensitivity (SI) every intervention– Measuring SI is a direct surrogate of patient response to all aspects of metabolism,

and is not available without a (good) model– Using just BG level is a very poor surrogate because it lacks insulin/nutrition context.

Like trying to estimate kidney function from just urine output – it lacks context

So, because we know the risk …

• We can minimise interventions, measurements and clinical effort with confidence and exact knowledge of the risk

• We know what to do when nutrition changes, and can change it directly if we require!

• So, what’s the glycemic target you ask? To what level do we control?– All we know is that level is bad and so is variability with about 1M opinions as to

what and how much…. – We, of course, have an answer… we think…

• Measures both level and variability

• We examined 3 “intermediate ranges” that most would think are not at all different!

• And 4 thresholds (50, 60, 70 and 80%) versus outcome (odds ratio)

cTIB = cumulative time in band: exposure (badness) over time

cTIB• 1700 patients from SPRINT and before

SPRINT, and both arms (high and low) of Glucontrol trial in 7 EU countries

• Is there a difference between 7 and 8 mmol/L or 3-4 mmol/L of variability???

• Yes, significantly so from day 2-3 onward

• Difference is more stark if you eliminate patients who have at least 1 hypo (BG < 2.2)

• We think the answer is clear and know how to safely achieve those goals

• Because you can calculate it in real time you can use it as an endpoint for a RCT

Day (1-14)

Surv

ival

Odd

s R

atio

4.0 – 7.0 5.0 – 8.0 4.0 – 8.0

cTIB > 50%

cTIB > 60%

cTIB > 70%

cTIB > 80%

“SPRINT”: Specialised Relative Insulin and Nutrition Tables

Chase JG, Shaw G, Le Compte A, Lonergan T, Willacy M, Wong XW, Lin J, Lotz T, Lee D, Hann C: Implementation and evaluation of the SPRINT protocol for tight glycaemic control in critically ill patients: a clinical practice change. Crit Care 2008, 12:R49

P=0.077 P=0.023 P=0.012 P=0.010P=0.244

LOS ≥ 2 days LOS ≥ 3 days LOS ≥ 4 days LOS ≥ 5 daysLOS ≥ 1 day

The horizontal blue line shows the mortality for the retro cohort. The green line is the total mortality of SPRINT patients against total number of patients treated on the protocol

Hospital mortality SPRINT/Pre-SPRINT

SOFA scores reduce faster with SPRINT and do so from day 2Organ failure free days: SPRINT = 41.6% > Retro = 36.6% (p<0.0001)Number of organ failures (% total possible) defined as SOFA > 2 for 1 SOFA

score component: SPRINT = 16% < Retro = 19% (p<0.0001)

Why? Better resolution of organ failure…

Chase JG, Pretty CG, Pfeifer L, Shaw GM, Preiser JC, Le Compte AJ, Lin J, Hewett D, Moorhead KT, Desaive T: Organ failure and tight glycemic control in the SPRINT study. Crit Care 2010, 14:R154.

At yesterday's cost…C

ost p

er a

nnum

Cos

t per

pat

ient

$0.5 M

$1.5M

Pre-SPRINT SPRINT

$2M

$1 M

Cos

t per

yea

r

Transfusions

Dialysis

Inotropes

Laboratory

Ventilation

Antimicrobials

Glucose control

ICU Costs

$0.5 M

$1.5M

Pre-SPRINTPre-SPRINT SPRINTSPRINT

$2M

$1 M

Cos

t per

yea

r

Transfusions

Dialysis

Inotropes

Laboratory

Ventilation

Antimicrobials

Glucose control

ICU Costs

Transfusions

Dialysis

Inotropes

Laboratory

Ventilation

Antimicrobials

Glucose control

ICU Costs

Pfeifer L, Chase JG, Shaw GM, “What are the benefits (or costs) of tight glycaemic control? A clinical analysis of the outcomes,” Univ of Otago, Christchurch, Summer Studentship 2010

In summary …

• We approach glycemic control like any problem– Understand the system (what happens when I do “x”?)– Understand the risk (how likely will the situation change? What happens if it does?)

• We accomplish this by using models– Of metabolism to understand the system– Of variability to understand the risk

• From understanding the system and understanding the risk we can dose to get safe and effective glycemic control by understanding that there are two ways (not just 1!) to lower (or raise) glycemia.

• STAR = Stochastic TARgeted glycemic control– Semi-automated– Reduced effort– Improved confidence and performance

A brief pause for reflection …

The future: digital human?

But beware of hyperbole!“Scientists have developed a technology that can bring people back from the dead up to seven hours after their hearts have stopped – and want it installed routinely in hospitals and even ambulances

“Ecmo (sic) machines, which act like heart bypass systems, but can be fitted in minutes are already used to save cardiac arrest victims in Japan and South Korea, where they are credited with reviving people long after they have apparently died

“ [Dr Sam] Parnia ...director of resuscitation at Stony Brook University...is publishing a book, The Lazarus Effect, about how death-reversing technologies are changing medicine”

The RCT methodology was created to validate responses to interventions amongst populations of highly complex biological systems (aka humans).

Prediction of individual responses is not possible because it requires an understanding beyond our current state of knowledge.

Clinical ‘trialists’ therefore must regard all patients as “black boxes”

State-of-the-art computing can be used model and validate these relationships; previously only guessed at, to create new knowledge and understanding.

Future RCTs should clinically validate interventions based on model-based therapeutics; a one-model-fits all approach.

(Patient-specific)

Acknowledgements Glycemia PG Researchers

Thomas LotzJess LinAaron LeCompte

Jason Wong et alHans Gschwendtner

Lusann

Yang

Amy Blakemore &

Piers Lawrence

Carmen Doran

Kate Moorhead Sheng-Hui WangSimone

Scheurle

Uli

Goltenbott Normy Razak Chris PrettyJackie

Parente

Darren Hewett James RevieFatanah Suhaimi

UmmuJamaludin

Leesa PfeiferHarry ChenSophie PenningStephan Schaller

Sam Sah PriBrianJuliussen Ulrike Pielmeier

Klaus Mayntzhusen Matt Signal Azlan Othman Liam Fisk Jenn Dickson

Math, Stats and Engineering Gurus

Dr Dom LeeDr Bob Broughton

Dr Paul Docherty

Prof Graeme Wake

The Danes

Prof Steen Andreassen

Dunedin

Dr Kirsten McAuley Prof Jim Mann

Acknowledgements Glycemia - 1

Geoff Shaw and Geoff Chase

Don’t let this happen to you!

Some guy named Geoff

The Belgians

Dr Thomas DesaiveDr Jean-Charles

Preiser

Hungarians

Dr Balazs Benyo

Belgium: Dr. Fabio Taccone, Dr JL Vincent, Dr P Massion, Dr R RadermeckerHungary: Dr B Fulesdi, Dr Z Benyo, Dr P Soos, Dr I Attila, and 12 others ...

... And all the clinical staff at over 12 different ICUs

Acknowledgements (Neonatal) Glycemia - 2

And Dr Adrienne Lynn and all the clinical staff at Christchurch Women's Hospital, and all the clinical staff Waikato Hospital

Prof Jane Harding Ms Deb Harris RN Dr Phil Weston

Auckland and Waikato

eTIME (Eng Tech and Innovation in Medicine) Consortia

4 countries, 7 universities, 12+ hospitals and ICUs and 35+ people

AcknowledgementsDept of Intensive Care