ai and ml for predicting covid-19

AI and ML for Predicting COVID-19

Malik Magdon-Ismail,Computer Science, Rensselaer.

Shout-Out:

Rensselear IDEA

J. Hendler, K. Bennet, J. Erickson, MANY good students.

Scales of COVID-19

World ∼ 8 billion

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 2 / 11 Two Faces →

Scales of COVID-19

USA ∼ 330 million


Scales of COVID-19

NY State ∼ 20 million


Scales of COVID-19

Albany/Troy/Cap Dist ∼ 1 million


Scales of COVID-19

Rensselaer ∼ 10 thousand

New Infections Over Previous 14 Days

# Students: 6806

Testing: every 7 days, 0% of students

Infections: 1.4%

Budget: 100000 tested

R0: 7.44

R(no test): 1.64

R(test): 1.64

Jan 24

Feb 13

Mar 5

Mar 25

Apr 14

0

1

5

10

20

Infe

ctio

n C

ount

(% o

f stu

dent

s)


Scales of COVID-19

Party at Rensselaer ∼ 20

Chances to Get COVID on 14-Feb-2021 (no masks)

20 40 60 80 100

Size of Event/Party

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5H

ours

at E

vent/P

art

y

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1


Scales of COVID-19

vaccines, virology, genomics


Two Sides of COVID Modeling

Epidemiological ModelingHarvard-model, Imperial-model, UW-model, Your-model, My-model, . . .

AI and Machine Learning PredictionWhat the data says vs. What we think ought to be.

Engineering success vs. Biological correctness.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 3 / 11 The Challenge →

The Race To Predict Ventilator Demand

NYC Capital District

Mar/02 Mar/10 Mar/18 Mar/260

1

4

9

16

25

36

Mar/04 Mar/10 Mar/16 Mar/220

5

10

15

20

25

30

35

40

45

Infection counts: very noisy dirty data.Predictions must be local: mobility patterns, density, social distancing, weather, . . . .

Smaller regions: more noisy; more sparse.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 4 / 11 A Easier Example →

A Easier Example

True “biological” law: quadratic growth.

Quadratic Fit + Extrapolate

1 2 3 4 5 60

10

20

30

40

50

60

70

80 Observed

True Quadratic Law

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 5 / 11 Regularization →

A Easier Example


Quadratic Fit + Extrapolate

1 2 3 4 5 60

10

20

30

40

50

60

70

80 Observed

True Quadratic Law

Quadratic Fit


A Easier Example


Quadratic Fit + Extrapolate Linear Fit + Extrapolate

1 2 3 4 5 60

10

20

30

40

50

60

70

80 Observed

True Quadratic Law

Quadratic Fit

1 2 3 4 5 60

10

20

30

40

50

60

70

80 Observed

True Quadratic Law

Linear Fit


A Easier Example


Quadratic Fit + Extrapolate Linear Fit + Extrapolate

1 2 3 4 5 60

10

20

30

40

50

60

70

80

1 2 3 4 5 60

10

20

30

40

50

60

70

80

Eout ≈ 34 Eout ≈ 14 ✓


A Stunning Nugget From The Theory of Learning

When there is noise,

Simpler can be better than correct.

1 2 3 4 5 60

10

20

30

40

50

60

70

80

1 2 3 4 5 60

10

20

30

40

50

60

70

80

What we would like to learn versus what we can learn.The data determines what we can learn

Harvard-model, Imperial-model, UW-model, Your-model, My-model, . . .

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 6 / 11 Let’s Predict →

A Stunning Nugget From The Theory of Learning

When there is noise,

Simpler can be better than correct.

1 2 3 4 5 60

10

20

30

40

50

60

70

80

1 2 3 4 5 60

10

20

30

40

50

60

70

80

What we would like to learn versus what we can learn.The data determines what we can learn

Harvard-model, Imperial-model, UW-model, Your-model, Simple–robust–adaptable model, . . .

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 6 / 11 Let’s Predict →

Let’s Predict For The Capital District

Mar/04 Apr/03 May/03 Jun/020

50

100

150

200

250

How quickly is it spreading?

How large is the pasture?

Capital District ∼ 1M.

Extrapolation is hard.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 7 / 11 Keep It Simple →



50

100

150

200

250





Disaster!




50

100

150

200

250

changepoint





Changepoints make it impossible.

Disaster!


Keep It Simple, Really Simple. But, Adaptive

U M

S

R

βMU/N

γ∆M(t − k)

(1 − γ)∆M(t − k)

U: Uninfected.

M: Contagious.

S: Symptomatic.

R: Recovered.

Parameters:N, β, α, γ.

Robust changepoints.

1 Robustly determine changepoints.

2 Robustly fit. Gray is uncertainty.

3 State persists across changepoints.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →


U M

S

R

βMU/N

γ∆M(t − k)

(1 − γ)∆M(t − k)

U: Uninfected.

M: Contagious.

S: Symptomatic.

R: Recovered.



1 Robustly determine changepoints.

2 Robustly fit. Gray is uncertainty.

3 State persists across changepoints.

How: Even simpler analytic model pre-calibrates.



U M

S

R

βMU/N

γ∆M(t − k)

(1 − γ)∆M(t − k)

U: Uninfected.

M: Contagious.

S: Symptomatic.

R: Recovered.



We get current state:

Infected and contagious. Immune. Social distancing.

Predictions assuming stabilized behavior.


COVID-War-Room https://covidwarroom.idea.rpi.edu

Capital District North Carolina

All US Counties. All Countries.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 9 / 11 COVID-Back-To-School →

COVID-Back-To-School https://covidspread.idea.rpi.edu

Who’s bringing covid to campus?

Ambient county infection rate?

COVID-War-RoomJan 19:

∼24 cases,∼20% immunity.

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 10 / 11 Tools to Policy →


Infection Growth from Start of Semester

# Students: 6806


Infections: 9.1%


R0: 7.44

R(no test): 1.64

R(test): 1.64

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)


# Students: 6806


Infections: 1.4%


R0: 7.44

R(no test): 1.64

R(test): 1.64

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)




# Students: 6806


Infections: 9.1%


R0: 7.44

R(no test): 1.64

R(test): 1.64

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)


# Students: 6806


Infections: 1.4%


R0: 7.44

R(no test): 1.64

R(test): 1.64

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)


20 40 60 80 100

Size of Event/Party

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

Hours

at E

vent/P

art

y

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1Chances to Get COVID on 14-Feb-2021 (masks)

20 40 60 80 100

Size of Event/Party

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

Hours

at E

vent/P

art

y

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1




# Students: 6806


Infections: 1.6%


R0: 7.44

R(no test): 1.06

R(test): 0.98

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)


# Students: 6806


Infections: 0.9%


R0: 7.44

R(no test): 1.06

R(test): 0.98

Jan 2

4

Feb 13

Mar

5

Mar 2

5

Apr 14

0

1

5

10

20

Infe

ction C

ount (%

of stu

dents

)


20 40 60 80 100

Size of Event/Party

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

Hours

at E

vent/P

art

y

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1Chances to Get COVID on 14-Feb-2021 (masks)

20 40 60 80 100

Size of Event/Party

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

Hours

at E

vent/P

art

y

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

Rensselaer: 1.5% ≈ 60. 18 infections so far.


Tools to Policy

We have tools to model spread at all scales.

In policy making, all scales are relevant. Decisions should take a holistic view.

The spread of COVID is just one factor that influences these decisions.

. . .

I really enjoyed giving this talk

Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 11 / 11

ai and ml for predicting covid-19

Documents