Possible factor (1)
0 and 1 are mandatory. 2 can be incorporated easily.
Are items 3 to 6 required? What are the strongest factors?
0. Past (current) transition in the number of new positive cases
1.Infectability of variants (this can be set upon making estimates from statistics, etc.)
2.Impact of vacations (three or more consecutive holidays, etc.)
3. Vaccination effect: Numeric fitting by comparing with the actual values
*Dedicated network (no change from the existing one)
4. Flows of people (DoCoMo mobile spatial statistical, Google Mobility, etc.)
5. Behavioral change Twitter (keywords: drinking party and karaoke), nighttime retention population
6. Weather conditions Temperature, humidity, and weather
Mutually related items:
Nighttime retention population (flows of people and behavioral change)
Weather (weather conditions and flows of people)
Analysis result:
They can be roughly expressed by flows of people (major transfer stations), Twitter (drinking party), vacation definition, and vaccination's effect. If limited to the short term (e.g. diffusion or end phase only) or season, the other parameters may work significantly.
Notes on possible factors
〇Reconsideration of the usability of the nighttime retention population
Periodic meeting of the Cabinet Secretariat (9/27) Report by Dr. Nakata
Even cases with high correlations are not necessarily stable
Report by Dr. Kurahashi (oral report)
The time of “nighttime retention population” with the highest correlation changes.
If parameters with a high estimate accuracy can be extracted, we should be able to learn something from them.
Which parameters have high reproducibility for about half a year?
Prediction system based on AI (deep learning)
Based on deep learning (LSTM model), directly predicted the number of new positive cases/number of seriously ill patients per day (one-week average value), etc. To identify the input parameters, things other than the future number of positive cases should be known.
The data volume in appearance can be increased by using standardized data instead of that of prefectures. Estimate based on non-linear regression (the number of parameters are virtually unlimited)
*The values can be calculated within several months to the extent that the accuracy can be secured
1. E. A. Rashed and A. Hirata, “Infectivity upsurge by COVID-19 viral variants in Japan: evidence from a deep learning modeling.” Int. J. Environ. Res. Public Health, 2021.
Item 1: Transition in virus variants
When a variant appeared (date when it exceeded 20%):
1st wave (B.1.1): 20/3/13, 2nd wave (B.1.1.284): 20/5/26,
3rd wave (B1.1.214): 20/10/7, 4th wave (B.1.1.7 α): 21/3/19,
5th wave (AY.29 δ): 21/6/23
Date when the occupancy of the variant exceeded 80%:
1st wave: 20/4/1, 2nd wave: 20/6/12, 3rd wave: 20/11/9,
4th wave: 21/4/26, 5th wave: 21/7/21
Defined the relative infectability (effective reproduction number) and calculated the coefficient from the proportion
Example of estimating the 4th and 5th waves in Tokyo and Osaka
(Learning up to April 15, 2021 and March 4, 2021 for Osaka, and estimate for the following period for Tokyo)
Learned with the estimate period and verified the effectiveness of the constructed network with input data of things other than the number of new positive case being known
Right axis: Standardized value of flows of people at stations, infectability of the virus, and Twitter (drinking party) data
*Prepared using the Twitter data provided from NTT DATA by Toyoda Lab, Institute of Industrial Science, The University of Tokyo
Example of estimating the 4th and 5th waves in Tokyo and Osaka
(Learned up to April 15, 2021 for Tokyo and March 4, 2021 for Osaka)
Learned with the estimate period with input data of things other than the number of new positive case being known. Searched input parameters with strong relationships with the number of new positive cases
The nighttime retention population is not necessarily an optimal indicator for its estimation even if it has a causal relationship with the number of new positive cases.
*The other input parameters are being verified
Right axis: Standardized value of flows of people at stations, infectability of the virus, and Twitter (drinking party) data
Relationship between the nighttime retention population (Shinbashi, Shibuya Center-Gai, Ginza, and Higashiginza) and Twitter (drinking party and karaoke)
(Upper: No averaging, Lower: One-week average)
The nighttime retention population at 21:00 is strongly correlated with “Drinking party” and the 7-day average value is strongly correlated with “Drinking party + Karaoke”
*Prepared using the Twitter data provided from NTT DATA by Toyoda Lab, Institute of Industrial Science, The University of Tokyo
Consideration and future
・Even if the nighttime retention population is large, it does not necessarily mean risky behaviors. Possibility of time series changes in risks (e.g. asymptomatic infection)
・The Twitter data “drinking party” is correlated with the nighttime retention population, so there is a possibility that it may work as a surrogate for multiple risky behaviors.
・Flows of people at major stations are not risky in themselves, but infection may spread.
・Vaccination effect: August and onward (past reports). The effectiveness of the effective value considering the vaccine can be confirmed for the nighttime retention population as well.
・Similar trend to that of Osaka. The other parameters are being verified
About the omicron variant: Can be simulated with the following items
Effectiveness of vaccination on the delta variant per capita × Effectiveness on the omicron variant
Effective reproduction number of the omicron variant in Japan. How about aggravation??
Johannesburg for up to 2020/10 and Gauteng for the following period (the number of infected people in Johannesburg was approximated as 35% of Gauteng). 2021/11/23 was excluded as a singular value
