A note of a virtual conference on AI & CoViD-19 - Section 3

April 3, 2020

Full Agenda: HAI Stanford

Full video: Youtube

Edit (8/4/2020) I add 4 more sections.

Session III: Tracking the Epidemic - 3:01:45

Based on the JAMA paper on 3 Mar 2020.
How Taiwan prepares for the next crisis starts with the Communicable Disease Control Act.
How Taiwan designed a center command center with three databases: Insurance stockpile systems, Infectious Disease Stat System, and media surveillance.
In the first 30 days, how Taiwan recognized the crisis, including boarding people from Wuhan and the command center.
Linked dataset for big data analytics between insurance and immigration.
Government stockpile: 44M surgical mask, 1.0M N95.
Diamond Princess ship story.
Mobile app to track inventories for masks.
Infrared temperature map.
Communication and Politics.
Challenge for Taiwan.

Pathengon Genomics and mathematical models
Why is undetected infection important? Asymptomatic or subclinical infection, limited testing capacity, restricted criteria for testing.
Consequence: spreading without notice, impacted large regions, longer time to elimination.
How to estimate: Mass testing, but expensive and time consuming; modeling; viral genomic data.
An example of a viral genomic data by time series with Wuhan Data.
Lead to Recommendation of face mask use by CDC

Started in late December. It is a challenge to create a visualization (health map) because of a lack of resources.
Solved by going to local news and China sources.
It needs an army of people to get data, not just by machine learning and web scraping. So people created a network of researchers and professionals for every single case around the globe > A risk and intervention.
An example of Baidu data about actions when you have detailed information.
An example of crowd source data from Thermia. Data connected to a medical device.
Symptom checkers with AI chatbot called Buoy. Other side, collect data to increase local surveillance.
Another flu near you to crowdsource data.
Another example with CoViD near you. 260 people reported symptoms.
Mining for social measure: to understand social distancing impact with collaboration with CDC, e.g. mapping data between policy and impact.

Flu background: Influenza and author story with connection with CDC
Flu prevalence data: Most target interest of CDC is % influenza like illness started in 1997. Data collected weekly. Why? Based on the severity pyramid.
How to forecast a season epidemic:
- Choose a type of model: mechanistic vs. statistical;
- For stat models, collects lots of data like Google search queries, Wikipedia page hits, thermometer sales, etc.
- For stat models, long history of training data to learn predictive relationships.
- Either type, should model revisions (backcasting!)
- Either type, prob forecasts are key; as are ENSEMBLEs.
How to not forecast a pandemic: SEASONAL epidemic
- Digital surveillance sensor > erratic model;
- Without sensors, ILI decreases.
Information about CDC flu forecast challenge about %ILC and hospitalization.
CMU effort in data challenge at https://delphi.cmu.edu/crowdcast/

Shared a story of Wuhan quarantine along with a speaker’s acknowledgement about the scale, and how she started to know about the China virus tracking app with a concern about a pervasive data collection which is not accepted in the US.
GPS is perhaps a solution with anonymous data for contact tracing apps, which allows us to find individuals contacted by an illed person. The challenge is highly restricted information of medical data.
Talked about a tracking app for the flu by using bluetooth protocols which are localized, so can deconstruct data. It allows us not to collect data of a person to do intervention.
Bluetooth strength allows us to know about the proximity of the person between two people.
Some other teams also work on this topic: COVID watch, TraceTogether, PrivateKit: Safe Paths, Colgi: Community Epidemiology in Actions.

Healthcare access and scalability with 50% world with no access to essential health service, ~30% of the US adults under-insured. ⅓ Americans self-diagnose online, ~15 minutes to capture information, diagnose and recommend treatment. COVID-19 is putting a huge stress on the medical system.
Problem of telehealth: it only solves some problems, importantly can NOT SCALE.
A proposal towards AI powered scalable health systems with a human in the loop, mobile first care.
A demo of how it looks in practice (chat-based application) with client side and doctor side.
An example of UI for Personalized diagnostic assessment.
Some peer-reviewed research about diagnosis with an example of question similarity with automated question and answering & feedback loop. But it is hard to verify because of lacking ground truth.

COVID-19 will change science and the way we do science.
10-20 years change drastically with data and hypothesis. There is a shift from output science, but science. Data of COVID-19 becomes available leading to excitement.
Problem is that there is too much data without a catalog which does not have a catalog.
The point to make: Open data is about more than disclosure. It must be fair, findable, accessible, interoperable, reusable. Nevertheless most of the data is not fair.
Example of a failure to use standard terms makes datasets often impossible to search. E.g. name of variables (age, Age, AGE, ‘Age, age (after birth), age (in years), age(y), age[y], age[years])
Another example of bad meta data from NCBI BioSample 73% of “Boolean” metadata values are not actually true or false; 26% of data can not be parsed into integers, 68% of metadata entries that supposed represent terms from standard vocabularies do not actually do so.
Why does it matter? To find data from relevant studies, integrate new results, re-explore existing data and verify published samples.
Solution: CEDAR approach to standardize Metadata with three step process.
An application of a use of CEDAR with project VODAN to help researchers learn about epidemic.
Online data will never be fair until we standardize metadata with templates<-controlled terms<-technology that makes it easy for investigators to annotate their datasets in standardized and searchable ways.

The psychological impact of COVID-19: uncertainty/ anxiety, unemployment and loneliness -> The path of well-being and mental health.
Unemployment and life satisfaction: job loss highest rate ever since the WWII
Digital social transition might be a solution for loneliness.
Introduction to do text analysis on twitter nationwide of the US. Scraping data on 26/2-26/3 in 2015 counties: Urban counties Washing your hands, panic buying, work from home, canceled events, wash hands; Educated counties: Government, healthcare, articles/ news, will be canceled, testing; Older counties: Trump and economic impact; Younger counties: wash hands; Voted for Trump: like the flu; Most negative sentiment: Scared, Economy, Trump;