AI and ML for Predicting COVID-19 Malik Magdon-Ismail, Computer - - PowerPoint PPT Presentation

▶

Jan 25, 2023 114 likes •449 views

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

SLIDE 1

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.

SLIDE 2

Scales of COVID-19 World ∼ 8 billion

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

SLIDE 3

Scales of COVID-19 USA ∼ 330 million

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

SLIDE 4

Scales of COVID-19 NY State ∼ 20 million

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

SLIDE 5

Scales of COVID-19 Albany/Troy/Cap Dist ∼ 1 million

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

SLIDE 6

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 1 5 10 20

Infection Count (% of students)

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

SLIDE 7

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 5

Hours at Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

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

SLIDE 8

Scales of COVID-19 vaccines, virology, genomics

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

SLIDE 9

Two Sides of COVID Modeling

Epidemiological Modeling

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

AI and Machine Learning Prediction

What 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 →

SLIDE 10

The Race To Predict Ventilator Demand

NYC Capital District

Mar/02 Mar/10 Mar/18 Mar/26 1 4 9 16 25 36 Mar/04 Mar/10 Mar/16 Mar/22 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 →

SLIDE 11

A Easier Example

True “biological” law: quadratic growth. Quadratic Fit + Extrapolate

1 2 3 4 5 6 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 →

SLIDE 12

A Easier Example

True “biological” law: quadratic growth. Quadratic Fit + Extrapolate

1 2 3 4 5 6 10 20 30 40 50 60 70 80

Observed True Quadratic Law Quadratic Fit

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

SLIDE 13

A Easier Example

True “biological” law: quadratic growth. Quadratic Fit + Extrapolate Linear Fit + Extrapolate

1 2 3 4 5 6 10 20 30 40 50 60 70 80

Observed True Quadratic Law Quadratic Fit

1 2 3 4 5 6 10 20 30 40 50 60 70 80

Observed True Quadratic Law Linear Fit

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

SLIDE 14

A Easier Example

True “biological” law: quadratic growth. Quadratic Fit + Extrapolate Linear Fit + Extrapolate

1 2 3 4 5 6 10 20 30 40 50 60 70 80 1 2 3 4 5 6 10 20 30 40 50 60 70 80

Eout ≈ 34 Eout ≈ 14 ✓

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

SLIDE 15

A Stunning Nugget From The Theory of Learning

When there is noise,

Simpler can be better than correct.

1 2 3 4 5 6 10 20 30 40 50 60 70 80 1 2 3 4 5 6 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 →

SLIDE 16

A Stunning Nugget From The Theory of Learning

When there is noise,

Simpler can be better than correct.

1 2 3 4 5 6 10 20 30 40 50 60 70 80 1 2 3 4 5 6 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 →

SLIDE 17

Let’s Predict For The Capital District

Mar/04 Apr/03 May/03 Jun/02 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 →

SLIDE 18

Let’s Predict For The Capital District

Mar/04 Apr/03 May/03 Jun/02 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 →

SLIDE 19

Let’s Predict For The Capital District

Mar/04 Apr/03 May/03 Jun/02 50 100 150 200 250

How quickly is it spreading? How large is the pasture?

Capital District ∼ 1M.

Extrapolation is hard. Disaster!

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

SLIDE 20

Let’s Predict For The Capital District

Mar/04 Apr/03 May/03 Jun/02 50 100 150 200 250 changepoint

How quickly is it spreading? How large is the pasture?

Capital District ∼ 1M.

Extrapolation is hard. Changepoints make it impossible. Disaster!

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

SLIDE 21

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.

Robustly fit. Gray is uncertainty.

State persists across changepoints.

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

SLIDE 22

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.

Robustly fit. Gray is uncertainty.

State persists across changepoints.

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

SLIDE 23

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.

Robustly fit. Gray is uncertainty.

State persists across changepoints.

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

SLIDE 24

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.

Robustly fit. Gray is uncertainty.

State persists across changepoints.

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

SLIDE 25

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.

Robustly fit. Gray is uncertainty.

State persists across changepoints. How: Even simpler analytic model pre-calibrates.

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

SLIDE 26

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.

We get current state:

Infected and contagious. Immune. Social distancing.

Predictions assuming stabilized behavior.

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

SLIDE 27

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 →

SLIDE 28

COVID-Back-To-School

https://covidspread.idea.rpi.edu

Who’s bringing covid to campus? Ambient county infection rate?

          

COVID-War-Room

Jan 19:

∼24 cases, ∼20% immunity.

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

SLIDE 29

COVID-Back-To-School

https://covidspread.idea.rpi.edu

Infection Growth from Start of Semester

# Students: 6806 Testing: every 7 days, 0% of students Infections: 9.1% 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 1 5 10 20

Infection Count (% of students) 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 1 5 10 20

Infection Count (% of students)

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

SLIDE 30

COVID-Back-To-School

https://covidspread.idea.rpi.edu

Infection Growth from Start of Semester

# Students: 6806 Testing: every 7 days, 0% of students Infections: 9.1% 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 1 5 10 20

Infection Count (% of students) 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 1 5 10 20

Infection Count (% of students) 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 5 Hours at Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

Chances 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 Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

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

SLIDE 31

COVID-Back-To-School

https://covidspread.idea.rpi.edu

Infection Growth from Start of Semester

# Students: 6806 Testing: every 7 days, 0% of students Infections: 9.1% 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 1 5 10 20

Infection Count (% of students) 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 1 5 10 20

Infection Count (% of students) 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 5 Hours at Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

Chances 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 Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

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

SLIDE 32

COVID-Back-To-School

https://covidspread.idea.rpi.edu

Infection Growth from Start of Semester

# Students: 6806 Testing: every 7 days, 20% of students Infections: 1.6% Budget: 100000 tested R0: 7.44 R(no test): 1.06 R(test): 0.98 Jan 24 Feb 13 Mar 5 Mar 25 Apr 14 1 5 10 20

Infection Count (% of students) New Infections Over Previous 14 Days

# Students: 6806 Testing: every 7 days, 20% of students Infections: 0.9% Budget: 100000 tested R0: 7.44 R(no test): 1.06 R(test): 0.98 Jan 24 Feb 13 Mar 5 Mar 25 Apr 14 1 5 10 20

Infection Count (% of students) 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 5 Hours at Event/Party

0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

Chances 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 Event/Party

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.

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

SLIDE 33

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