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

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

Scales of COVID-19 NY State ∼ 20 million Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 2 / 11 Two Faces →

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 →

Scales of COVID-19 Rensselaer ∼ 10 thousand New Infections Over Previous 14 Days 20 # Students: 6806 Testing: every 7 days, 0% of students Infections: 1.4% 10 Budget: 100000 tested Infection Count (% of students) 5 R0: 7.44 R(no test): 1.64 R(test): 1.64 1 0 Jan 24 Feb 13 Mar 5 Mar 25 Apr 14 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 2 / 11 Two Faces →

Scales of COVID-19 Party at Rensselaer ∼ 20 Chances to Get COVID on 14-Feb-2021 (no masks) 0.1 5 0.09 4.5 0.08 4 0.07 3.5 Hours at Event/Party 0.06 3 0.05 2.5 0.04 2 0.03 1.5 0.02 1 0.01 0.5 0 20 40 60 80 100 Size of Event/Party Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 2 / 11 Two Faces →

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

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 →

The Race To Predict Ventilator Demand NYC Capital District 36 45 40 25 35 16 30 25 9 20 4 15 10 1 5 0 0 Mar/02 Mar/10 Mar/18 Mar/26 Mar/04 Mar/10 Mar/16 Mar/22 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 Observed 80 True Quadratic Law 70 60 50 40 30 20 10 0 1 2 3 4 5 6 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 5 / 11 Regularization →

A Easier Example True “biological” law: quadratic growth. Quadratic Fit + Extrapolate Observed 80 True Quadratic Law Quadratic Fit 70 60 50 40 30 20 10 0 1 2 3 4 5 6 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 5 / 11 Regularization →

A Easier Example True “biological” law: quadratic growth. Quadratic Fit + Extrapolate Linear Fit + Extrapolate Observed Observed 80 80 True Quadratic Law True Quadratic Law Quadratic Fit Linear Fit 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 1 2 3 4 5 6 1 2 3 4 5 6 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 5 / 11 Regularization →

A Easier Example True “biological” law: quadratic growth. Quadratic Fit + Extrapolate Linear Fit + Extrapolate 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 1 2 3 4 5 6 1 2 3 4 5 6 E out ≈ 34 E out ≈ 14 ✓ Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 5 / 11 Regularization →

A Stunning Nugget From The Theory of Learning When there is noise, Simpler can be better than correct. 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 1 2 3 4 5 6 1 2 3 4 5 6 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. 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 1 2 3 4 5 6 1 2 3 4 5 6 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 250 How quickly is it spreading? 200 How large is the pasture? 150 Capital District ∼ 1M. 100 50 0 Mar/04 Apr/03 May/03 Jun/02 Extrapolation is hard. Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 7 / 11 Keep It Simple →

Let’s Predict For The Capital District 250 How quickly is it spreading? 200 How large is the pasture? 150 Capital District ∼ 1M. 100 50 0 Mar/04 Apr/03 May/03 Jun/02 Extrapolation is hard. Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 7 / 11 Keep It Simple →

Let’s Predict For The Capital District 250 How quickly is it spreading? 200 How large is the pasture? 150 Capital District ∼ 1M. 100 50 0 Mar/04 Apr/03 May/03 Jun/02 Extrapolation is hard. Disaster! Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 7 / 11 Keep It Simple →

Let’s Predict For The Capital District 250 How quickly is it spreading? 200 How large is the pasture? 150 Capital District ∼ 1M. 100 50 0 Mar/04 Apr/03 May/03 Jun/02 changepoint 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 →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R U: Uninfected. M: Contagious. S: Symptomatic. R: Recovered. Parameters: N, β, α, γ . Robust changepoints. Robustly determine changepoints. 1 Robustly fit. Gray is uncertainty. 2 State persists across changepoints. 3 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R U: Uninfected. M: Contagious. S: Symptomatic. R: Recovered. Parameters: N, β, α, γ . Robust changepoints. Robustly determine changepoints. 1 Robustly fit. Gray is uncertainty. 2 State persists across changepoints. 3 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R U: Uninfected. M: Contagious. S: Symptomatic. R: Recovered. Parameters: N, β, α, γ . Robust changepoints. Robustly determine changepoints. 1 Robustly fit. Gray is uncertainty. 2 State persists across changepoints. 3 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R U: Uninfected. M: Contagious. S: Symptomatic. R: Recovered. Parameters: N, β, α, γ . Robust changepoints. Robustly determine changepoints. 1 Robustly fit. Gray is uncertainty. 2 State persists across changepoints. 3 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R U: Uninfected. M: Contagious. S: Symptomatic. R: Recovered. Parameters: N, β, α, γ . Robust changepoints. Robustly determine changepoints. 1 Robustly fit. Gray is uncertainty. How: Even simpler analytic model pre-calibrates. 2 State persists across changepoints. 3 Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 8 / 11 COVID-War-Room →

Keep It Simple, Really Simple. But, Adaptive γ ∆ M ( t − k ) S βMU/N U M (1 − γ )∆ M ( t − k ) R 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 →

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 COVID-War-Room Who’s bringing covid to campus?     Jan 19:   ∼ 24 cases,  Ambient county infection rate?    ∼ 20% immunity.  Creator: M. Magdon-Ismail, November 12, 2020 AI/ML for COVID-19: 10 / 11 Tools to Policy →