Archimedes Model Introduction

A KAISER PERMANENTE INNOVATION

An Overview of the Archimedes Healthcare Model

The Archimedes ProjectKaiser Permanente


This presentation will address three main topics:

• What the Archimedes model is

• How it can be used

• How we know it works


Until now, there have been four main kinds of mathematical models in health care

• Biological modeling– e.g. glucose metabolism, Starling heart equation

• Clinical Medicine– e.g. Regression equations, Markov models

• Operations research / management science– e.g. Schedule operating rooms, plan hospitals

• Economic / system resources– e.g. Forecasting


Archimedes is a new type of mathematical model of health care:

it includes all four components

Clinical medicine

Biological modeling

Care processes

System resources


Archimedes is also new because it is anchored to reality

through clinical trials

Clinical medicine

Biological modeling

Care processes

System resources


This presentation will address four main topics. Let’s begin with:

• What the Archimedes model is• How it can be used• How we know it works• What we plan to do with it• How you will be able to get access to it• What’s on this website


Archimedes is

• Object oriented

• Continuous

• Physiology based

• Very deep and very broad

• and it operates at the level of detail that clinicians and administrators consider essential for making decisions


Now let’s define those terms. This is what we mean by “object oriented”

• All the important objects in reality have corresponding objects in the model, one-to-one, at a high level of detail

Patients, organs, parts of organs

Facilities

Health care personnel

Equipment, supplies

Policies and procedures

Budgets, regulations, more


More specifically, this means that

• There are thousands of simulated people in the model, each one of them different

• They can get sick, and go see simulated doctors• In simulated offices• And get simulated tests, that use simulated

pieces of equipment• And get simulated treatments• And so forth


This is what we mean by “continuous”

• Biological variables that are continuous in reality are continuous in the model (e.g. BP, LDL cholesterol)– No “clinical states”, “strata”

• Time is continuous; any event can occur at any time– No “ticks”, “steps” or “annual jumps”


The concept of “physiology based”needs more explanation. And we’re

going to need a volunteerI’ll

volunteer

Joe. Joe Miner

Thanks. What’s your name?

OK Joe. I hope you’re not

ticklish…


… because we’re going to need to look inside

your belly.Can you come a bit

closer?

Sure

A little bit closer, if you can

How’s this?

ThatThatThat’’’s good.s good.s good.Nice belly, Joe.Nice belly, Joe.Nice belly, Joe.

Now weNow weNow we’’’re going to need to re going to need to re going to need to do a sort of xdo a sort of xdo a sort of x---rayrayray

OK.OK.The reason weThe reason we’’re re

doing this is that we doing this is that we want to show you that want to show you that

every simulated every simulated person in the model, person in the model, like Joe, has organs, like Joe, has organs,

such assuch as……

PancreasPancreasPancreasCirculatory Circulatory Circulatory systemsystemsystem

HeartHeartHeart

LiverLiverLiver

Fat cellsFat cellsFat cells

GutGutGut MuscleMuscleMuscle

And every organ has all of its types of cells, like these beta

cells of the pancreas

The organs and The organs and their cells carry out their cells carry out

metabolic metabolic functions,functions,such assuch as……

GLUCOSE UPTAKE BY FAT

GLUCOSE UPTAKE BY MUSCLE

SUGARS

HEPATIC GLUCOSE PRODUCTION

GLYCOGENGLYCOGENGLYCOGEN

GLUCOSEGLUCOSEGLUCOSE

CIRCULATING GLUCOSE

In the model, these In the model, these functions are regulated just functions are regulated just as they are in a real person. as they are in a real person. For example, in response to For example, in response to

the circulating levels of the circulating levels of glucose, the pancreatic beta glucose, the pancreatic beta

cells will secrete insulincells will secrete insulin

GLYCOGEN


GLUCOSE UPTAKE BY MUSCLESUGARS


GLUCOSE INSULIN

CIRCULATING GLUCOSE

And diseases can occur. And diseases can occur. For example, in a person For example, in a person with diabetes the effect of with diabetes the effect of

insulin on uptake of insulin on uptake of glucose by the muscle glucose by the muscle and fat is decreased.and fat is decreased.

GLYCOGEN




GLUCOSE INSULIN

Diabetes

Diabetes

Diabetes also affects the Diabetes also affects the effect of insulin on effect of insulin on

production of glucose by production of glucose by the liver cellsthe liver cells

GLYCOGEN




GLUCOSE INSULIN

Diabetes

IfIf we were to illustrate the we were to illustrate the biological variables and biological variables and

relationships in the model relationships in the model that affect the metabolism that affect the metabolism of glucose, it would look of glucose, it would look

something like thissomething like this

Type 1Diabetesfeature

Coronaryartery

stenosis

Race/ethnicity

Sex

Age

Gliburide

LDLcholesterol

Pulse \pressure

Arterialcompliance

Meanarterial

pressure

OGT

Diabetesdiagnosis

BMI

Height

Cardiacoutput

To theNeuropathy

model

To theNephropath

y model

To theCoronary

arterydiseasemodel

To theRetinopathy

model

Hypo-glycemia

Familyhistory

Diabetescardiac risk

factor

FPG

Randomplasmaglucose

OGTT test

Propensityto fatigue

Propensityto polyuria

Thirst

Fatigue

Polyuria

Perception

Memory

Patienttakes action

Blurredvision

Propensityto thirst

Propensityto blurred

vision

FPG

Diabetesblood

pressurefactor

Insulintreatment

Systolicblood

pressure

Peripheralresistance

Untreatedinsulin level

Insulinefficiency(Muscle)

Smoking

FPG

Randomplasma

glucose test

HbA1c test

FPG testRandomerror andvariation

Totreatmentmodels

Careprocesses

Triglycerides

Age, sex,race/

ethnicity


Metformin

Diet andexercise

Weight

HDLcholesterol

Insulinproduction(Pancreas)

Unexplainedvariance in

OGT

Keto-acidosis

Fractionalchange in

Insulin

Urineketone test

Glucoseproduction

by liver

Normal liverglucose

production

Insulinefficiency

(Liver)

JointDiabetesfeature

Insulin level

Glucoseuptake by

muscle




Coronaryartery

stenosis

Race/ethnicity

Sex

Age

Gliburide

LDLcholesterol

Pulse \pressure

Arterialcompliance

Meanarterial

pressure

OGT

Diabetesdiagnosis

BMI

Height

Cardiacoutput

To theNeuropathy

model

To theNephropath

y model

To theCoronary

arterydiseasemodel

To theRetinopathy

model

Hypo-glycemia

Familyhistory


factor

FPG

Randomplasmaglucose

OGTT test



Thirst

Fatigue

Polyuria

Perception

Memory

Patienttakes action

Blurredvision

Propensityto thirst


vision

FPG

Diabetesblood

pressurefactor

Insulintreatment

Systolicblood

pressure




Smoking

FPG

Randomplasma

glucose test

HbA1c test


Totreatmentmodels

Careprocesses

Triglycerides

Age, sex,race/

ethnicity


Metformin

Diet andexercise

Weight

HDLcholesterol



OGT

UKPDSdata

Keto-acidosis

Fractionalchange in

Insulin

Urineketone test

Glucoseproduction

by liver

Normal liverglucose

production

Insulinefficiency

(Liver)


Insulin level

Glucoseuptake by

muscle

All of that, and many, many more variables that are important to the complications of diabetes and to other diseases, are being calculated continuously in every simulated person in the model


And there are lots of simulated people, thousands in fact,

all of them with different risk factors,

physiologies, behaviors, et cetera.(The differences are shown as different colors of their

aprons)


Thanks Joe, you can go back to the

model now.You’re

welcome.Good bye


That was “physiology-based”. Let’s move on now.

Here’s what we mean by “broad and deep”• Object oriented

• Continuous

• Physiology based

• Very deep and very broad

• and it operates at the level of detail that clinicians and administrators consider essential for making decisions


“Very broad and very deep ” means that the model includes

• Not only– Anatomy, Physiology, Pathology, Signs and

symptoms, Tests and treatments

• But also– Patient and provider behaviors – Care processes: e.g. guidelines, disease

management, CQI– System resources: e.g. facilities, personnel,

equipment, supplies, costs …


To get the idea of “very deep and very broad”, let’s follow a problem from the

cellular level to the parking lot• All the simulated people in the model have all the

important organs. E.g. hearts, kidneys, immune systems…

• All the simulated organs have all the important parts. E.g. hearts have coronary arteries, ventricles, myocardium, sino-atrial node…

• All the part have all the important subparts. e.g. arteries have lumens, walls…

• All the organs and their parts have functions. e.g. arteries carry blood to the myocardium


There’s more

• Things can go wrong with the organs or their functions. e.g. coronary arteries can occlude.

• When something goes wrong, it affects other things. e.g. when arteries occlude, blood flow to the myocardium is reduced. When blood flow is reduced, the feels pain (angina).

• All the simulated patients have behaviors; when the pain reaches a threshold, the patient makes a simulated call to a simulated hospital.

• Simulated telephone operators at simulated call centers use simulated protocols to triage calls.


And still more• Patients go to simulated emergency

departments. (This is where the parking lot comes in)

• They are seen by simulated personnel (e.g., nurses).

• Simulated tests are performed…tests use equipment and supplies

• Diagnoses are made…and mistakes are made

• Patients are admitted, rooms are occupied, …• Treatments are given…


You get the idea. “Very broad” and “very deep” means that the effects of any encounter keep ramifying through

the simulated health care system, just as happens in the real world

• Logistics happen…• Utilization occurs…• Outcomes occur…• Costs occur…


There’s one more thing to discuss,the level of detail

• Object oriented• Continuous• Physiology based • Very deep and very broad•• It operates at the level of detail that It operates at the level of detail that

clinicians and administrators consider clinicians and administrators consider essential for making decisionsessential for making decisions


The level of detail is determined by the people who will use the model

• If a clinician says that a particular variable is critical to their own thinking, and the model won’t be credible without it …

• We include the variable– e.g. ejection fraction, insulin level, “beta cell

fatigue”, peripheral resistance• Ditto for administrators

– E.g. there are 37 different types of office visits


Now let’s talk about:

• What the Archimedes model is• How it can be used• How we know it works


The purpose of Archimedes is to create a virtual world

• Virtual people • who have virtual physiologies• get virtual diseases• have virtual signs and symptoms• go to virtual doctors• get virtual tests and treatments• and have virtual outcomes• just like real people


• Guidelines• Disease management programs• Nursing algorithms• Changes in care processes• Continuous quality improvement programs• Performance measures and targets

If it works, then we can use the virtual world to try out, evaluate, explore, optimize, plan,

improve, predict…lots of different things.Such as…


• Priority setting• Strategic goals• Research• How “ideal” research trials translate to

“real” clinical settings• Logistics, use of resources• Costs, cost-effectiveness, and value• Planning

And more…


The idea is that we can use the virtual world to learn the effects of lots of different types of

interventions that would be infeasible to study through empirical methods (more research), for a

variety of reasons:• Too high a cost• Too long time needed for a new study• Too many patients needed• Unwillingness of patients and/or physicians to

participate• Too large a number of options to study• Technologies that are changing too rapidly


This is very important, because

• An empirical study – e.g. an evaluation study or clinical trial – of a single intervention will – cost hundreds of thousands, or even

hundreds of millions of dollars– require hundreds, or thousands of people– take years

• Whereas an Archimedes simulation takes much less time and costs less


Sounds good. And I’m happy to help out. But if you don’t mind, I’ve got a question.All this is premised on an assumption that it works

right. How do you know that it works?

I mean, I know that my

equations are working fine. But I’m not so certain about the other

jokers who are in here with me.


Thanks Joe, that is indeed a crucial question:

• What the Archimedes model is• How it can be used• How do we know it works?


Here’s how:

• We compare what happens in the virtual world with what happens in the real world,

• in lot’s of different settings,• that test the parts of the model that will be

used for the types of applications we want to do– i.e. similar people, similar treatments, similar

outcomes


The trial is conducted

They get the treatments specified in the protocols of the trial

A trial involves people who meet the entry criteria for the trial

And the outcomes are measured

The best way to do this is to use The best way to do this is to use clinical trials because we know the clinical trials because we know the

most about themmost about them


Real clinicaltrial

Treatments

People RealOutcomes

We count the outcomes, using the same definitions and protocols

We use the same criteria to select people

We have simulated physicians follow the same treatment protocols

And we compare the results

We use the virtual world of the model to We use the virtual world of the model to simulate the realsimulate the real--world trialworld trial

We let the model do its thing


There are 10 steps involved in performing a validation

1. We start with a large population of simulated people, just like a simulated city

2. We identify the inclusion criteria for the trial3. We randomly select from the large

population a sample who meet the inclusion criteria for the trial

4. Confirm that major characteristics match (“Table 1”)

5. Randomize the simulated people


There are 10 steps involved in performing a validation (Cont’d)

6. Identify the treatment protocols used in the trial

7. Have simulated providers apply the treatment protocols to the simulated people

8. Have the simulated providers apply the follow-up protocols

9. Record the results seen in the real trial10. Compare the simulated results to the real

results


You can think of these validations as being previews of real applications• We used the populations that were in the

trials– But we could have used the population that

would be candidates for your guideline (or performance measure, etc)

• Similarly for the tests, treatments, outcomes, etc.


VA-HIT

IDNTUKPDSHPS

LewisDPPWOSCOPS

IRMALIPIDLLRC

MICRO-HOPECAREHHS

DCCT secondary4-SSHEP

DCCT primaryHOPEMRC

Archimedes has been validated Archimedes has been validated against these trials, so faragainst these trials, so far


Here’s an example: the Heart protection Study

• Population: adults age 40 – 80, at high risk of MI because of high LDL, or other arterial occlusive disease, or diabetes

• Treatments: Simvastatin vs. placebo

• Size: 20,500

• Duration: 6 years


Major Coronary Events in Heart Protection Study

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5Years

Frac

tion

of p

atie

nts

Placebo group

Treated group

Solid lines are the real results

This was the rate of major coronary artery This was the rate of major coronary artery events in the placebo and treated groupsevents in the placebo and treated groups


The dotted lines showed what the The dotted lines showed what the model calculatedmodel calculated

Major Coronary Events in Heart Protection Study

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5Years

Frac

tion

of p

atie

nts

Dotted lines are the model’s results

Placebo group

Treated group


Here’s another example, the Diabetes Prevention Program

• Population: men and women over age 25. broad range of racial and ethnic groups. Prediabetes (IGT or IFG)

• Treatments: intensive lifestyle vs. Metformin vs. standard care

• Size: 3000• Mean duration: 2.8 years


DPP: Diabetes Progression

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0 1 2 3 4 5Time (years)

Frac

tion

ControlMetforminLifestyle

In this case, the results were calculated by In this case, the results were calculated by the model the model beforebefore the real results were the real results were

known. This is what the model predicted.known. This is what the model predicted.


The dotted lines are what the model The dotted lines are what the model predicted. The solid lines are the results predicted. The solid lines are the results

that were eventually published.that were eventually published.DPP: Diabetes Progression

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

0 1 2 3 4 5Time (years)

Frac

tion

ControlMetforminLifestyle


Here’s an example of a long term trial that measured the effects of treatment on

complications of diabetes: UKPDS• Population: middle aged men and women, with

newly diagnosed type 2 diabetes

• Treatments: intensive treatment vs. conventional treatment

• Size: 3900

• Duration: 15 years


UKPDS: MI (fatal and non-fatal)

0

0.05

0.1

0.15

0.2

0.25

0.3

0 2 4 6 8 10 12 14Time (years)

Frac

tion

of p

atie

nts

Conventional treatment

Intensive treatment

The solid lines are the real results

This shows the effect of treatment on the This shows the effect of treatment on the most important outcome: heart attacksmost important outcome: heart attacks


UKPDS: MI (fatal and non-fatal)

0

0.05

0.1

0.15

0.2

0.25

0.3

0 2 4 6 8 10 12 14Time (years)

Frac

tion

of p

atie

nts

The dotted lines are the model’s results

This is what was calculated by the modelThis is what was calculated by the model

Conventional treatment

Intensive treatment


0

0.1

0.2

0.3

0.4

0.5

0.6

0 0.1 0.2 0.3 0.4 0.5 0.6Results from trials

Res

ults

cal

cula

ted

by m

odel

We have done this for 74 different combinations We have done this for 74 different combinations of populations, treatments and outcomes. This of populations, treatments and outcomes. This chart compares results calculated by the model chart compares results calculated by the model (y(y--axis) against results of the real trial (xaxis) against results of the real trial (x--axis). axis).

Each dot represents an arm of a trial. The real result of the trial is plotted against the result calculated by the model. Perfect erfect correspondence would correspondence would be the 45be the 45ºº lineline


One would not expect perfect correspondence because of random factors, related to the number of people in the trial

• 71 of 74 are well within sampling error• The other 3 have good explanations

– either it just missed (e.g. p = .04) (which is to be expected statistically)

– or the description of the trial was incomplete • 54 of 74 are within ± 1 standard deviation• For all 74 exercises: r = 0.99


For some of the trials, some of the data from the trial were used to help build

parts of the model. The other trials are 100% independent

• For the exercises that were not used at all to build the model (100% independent), the correlation between the model and the trial was still extremely high

• The correlation was still r = 0.99


These validations are important for several reasons

• They give us confidence that for applications that are spanned by validations, the model is accurate

• They illustrate the types of applications that the model can do

• No other model of clinical medicine has been validated in this way

• The alternative to the model, expert judgment, has not been validated in this way either (no offense intended)


The “span” of the validations is important, because it indicates the types of populations,

organ systems, and treatments we can be confident about


Imagine that the disease that we want to understand is represented by an island.

We need to completely map out and reach every part of that island


Imagine that the available research can be represented as radio beacons on the island.

The bigger and better the trial, the wider its signal.


Now suppose that Archimedes has been validated against all of those trials.

Using this visual image, we can talk about the usefulness of Archimedes for new applications


If you ask it a question about this part of the island (A), it should do extremely well

A


If you ask it a question about this part of the island (B), it should do very well

B


If you ask it a question about this part of the island (C), it should do fairly well

C


If you ask it a question about this part of the island (C), it should do fairly well,

but we’d like to see a new trial done down here

C


If you ask it a question about this part of the island (D), we’ll say “OK, but only use the

model for “what if” type questionsWe’d rather help you design a new trial

D


If you ask it a question about that is out in the ocean (E), We’ll say “Archimedes can’t do that yet. Help us raise some money and

we’ll build a new model for that”

E


With that in mind, it is important that Archimedes has been validated for a wide

range of populationsAverage riskHypercholesterolemiaDiastolic hypertensionSystolic hypertensionMany risk factorsAngina or MIDiabetesMI average CholesterolMI high CholesterolPre diabetes

Newly diagnosed diabetesType 1, uncomplicatedType 1, retinopathyType 1, moderate to severe nephropathyType 2, uncomplicatedType 2, mild nephropathyType 2, moderate nephropathyYoung, Middle age, Old


…and organ systems

• Liver• Pancreas• Heart• Vascular• Nerves

• Muscle• Fat• Kidneys• Eyes• Lungs• Gut


…and treatments

• Nothing (placebo)• Cholestyramine• Gemfibrozil• Anti-hypertensive• Pravastatin• Simvastatin• ACE Inhibitors

• Angiotensin –II-receptor antagonists

• Insulin• Metformin• Sulphonylurea • General diet advice• Intensive lifestyle


That’s what we mean when we say that Archimedes is anchored to reality

Clinical medicine

Biological modeling

Care processes

System resources

(74 anchors)


What does all this mean?

• It is possible to write equations that represent human physiology, disease, treatments and outcomes using existing information

• In the “virtual world” created by these equations, simulated patients behave like patients in the real world, …

• …at least as we know it through the most important epidemiological studies and clinical trials


It means the virtual world can be used for many purposes, such as

• Guidelines• Performance measures• Disease management programs• Continuous quality improvement• Strategic goals and priority setting• Research planning• Estimating patient-specific outcomes


And that’s good, because it’s a whole lot easier to explore problems and

programs in the virtual world

• Much faster

• Much less expensive

• Much more flexible

• Can explore many more options

• Can tell you the critical points to watch


If you want to use the model, there is one more thing you will need to understand • Real people like you will still have to make the

final decisions • Archimedes will give you much better

information than you’ve ever had before • But there are always elements of a decision that

can not be quantified and that require value judgments

• This is the role of humans• Thus the model is a just a tool -- a very powerful

tool, but just a tool



Coronaryartery

stenosis

Race/ethnicity

Sex

Age

Gliburide

LDLcholesterol

Pulse \pressure

Arterialcompliance

Meanarterial

pressure

OGT

Diabetesdiagnosis

BMI

Height

Cardiacoutput

To theNeuropathy

model

To theNephropath

y model

To theCoronary

arterydiseasemodel

To theRetinopathy

model

Hypo-glycemia

Familyhistory


factor

FPG

Randomplasmaglucose

OGTT test



Thirst

Fatigue

Polyuria

Perception

Memory

Patienttakes action

Blurredvision

Propensityto thirst


vision

FPG

Diabetesblood

pressurefactor

Insulintreatment

Systolicblood

pressure




Smoking

FPG

Randomplasma

glucose test

HbA1c test


Totreatmentmodels

Careprocesses

Triglycerides

Age, sex,race/

ethnicity


Metformin

Diet andexercise

Weight

HDLcholesterol



OGT

UKPDSdata

Keto-acidosis

Fractionalchange in

Insulin

Urineketone test

Glucoseproduction

by liver

Normal liverglucose

production

Insulinefficiency

(Liver)


Insulin level

Glucoseuptake by

muscle

But watch out. I’m

gaining on you.

Documents

Archimedes Model Introduction