Changes in new positive cases in Jul-Sep period

Characteristics: Directly predict the numbers of new positive cases and seriously ill patients per day up to one month later considering flows and people and seasonality based on the deep learning model (LSTM model) (one-week average value)

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 is virtually unlimited)

Repeat learning every week to update the prediction model.
If flows of people are grasped, the daily number of newly infected people for approximately one month from now can be predicted (one-week average value)
(two-week average 20%): Up to the 4th wave
High accuracy up to two weeks later because the value of past flows of people of about two weeks is used

The initial value may change because the data of six prefectures are learned.
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.

The original variant of the novel coronavirus can be predicted at an average accuracy of 81.6% based on flows of people. Used Google mobility for flows of people. Non-linear regression of whole flows of people based on AI is the surrogate. For Tokyo, major information (transfer stations) alone is not enough.

１．Predict the number of new positive cases from the Japanese situation considering the impact of the variant (no definition of the vaccination rate)
２．Learned and corrected the data of other countries to consider the impact of vaccination in more detail

Model (1)：The impact of vaccination is not defined explicitly
・The input values include the weather information (minimum/maximum and average temperature), data on flows of people, weekdays or holidays, presence or non-presence of declaration of a state of emergency, and number of new positive cases to date as well as the newly added variation label (0: standard, 1: alpha, 2: delta)
・The future values of weather data and the data of flows of people are assumed to be similar to those of last year
・The applicable city of learning data is Tokyo
Model (2)
・Separated the prediction model to make it possible to consider only the impact of vaccination
・The input values of the network are the number of new positive cases and the effectiveness of vaccination
・The applicable cities of learning data are Tokyo and Tel Aviv
・The age composition is not considered (no data of Tel Aviv)

Effectiveness on the people flow increase-decrease rate at a transfer station as a surrogate (average value of 6 to 13 days before the corresponding day)
3rd wave (original variant)
4th wave (alpha variant)
5th wave (delta variant)
Based on these divisions of waves, the effective reproduction number is correlated with the increase-decrease rate of people flow at the major station.

*The effective reproduction number is excerpted from Toyo Keizai Online
*The 5th wave is an approximated straight line excluding the impact of the long holiday season
*The approximated straight line is based on the infection spread period

When people flow decreases, the effective reproduction number also decreases.
The four consecutive holidays in July and Obon (in August) greatly change the effective reproduction number.
3rd wave (original variant)
4th wave (alpha variant)
5th wave (delta variant)
Based on these divisions of waves, the effective reproduction number is strongly correlated with the increase-decrease rate of people flow at the major station (it works as a surrogate in Osaka).

*The effective reproduction number is excerpted from Toyo Keizai Online
*People flow is the average value of 6 to 13 days before the corresponding day
*The 5th wave is an approximated straight line excluding the impact of the long holiday season

Corrected the effective reproduction numbers based on the approximated straight lines of the number of new positive cases in Tokyo and Osaka
Estimate close to the normal impression.

*People flow is the average value of 6 to 13 days before the corresponding day
*The correction period of the effective reproduction number is defined as the period including the impact of flows of people during the consecutive holidays

○The estimated value by machine learning is almost the same as the number of new positive cases of the corrected value based on the approximated straight line.
The rises in the effective reproduction number associated with activities during the consecutive holidays in July and Obon (August) are more than predicted.
○Difference in prediction by machine learning in and after mid-September
As of July 22, the effect of vaccination was not estimated correctly.
Possible reason: Not only the type of the vaccine but also the increase-decrease rate at which the effective reproduction number becomes zero are different between Israel and Japan.
Infection prevention effect of the vaccine on the delta variant. Machine learning: Used the data of Tokyo (up to July 22) and Israel. Combination of models (1) and (2). For the vaccination effect, used the actual data of Tokyo.

Only the efficiency on the number of vaccinated people was not fixed
Prediction result in the case that the effectiveness of the prevention effect of vaccination is kept the same as that of July 22 (0.15 to the total population) (Learning period: From 2021/7/22)

If vaccination had been insufficient, it might have been about 10,000 and the period up to peaking-out might have been longer. The infection-prevention effect of vaccination will be clarified in and after mid-August.
Machine learning: Generate virtual data after learning by the end of September

Tokyo: Prediction of the number of new positive cases (weekly average) if weather conditions and flows of people are at the same level as last year. Considering the impact of vaccines with different networks

The start of the decrease was two weeks to one month earlier than predicted.
The result of predicting the decreasing speed did not change greatly.
The machine may have learnt past periodicity.

Osaka: Prediction of the number of new positive cases (weekly average) if weather conditions and flows of people are at the same level as last year. Considering the impact of vaccines with different networks

The decrease started about two weeks earlier than predicted. 　
The result of predicting the decreasing speed (inclination of the change) decreased greatly.
Cause (consideration): The value to be the threshold of the decrease in flows of people is different from that of Israel.

・The effective reproduction number increased due to the behaviors during the four consecutive holidays in July and Obon (August) (especially during vacations, there is a non-linear relationship with the flows of people).
・Estimate by machine learning (type of AI): If vaccination had been insufficient, the daily number of new positive cases (one-week average) might have been 10,000 (95% value 6,000–14,000) and the peak timing might have been around September 10.
・The decrease in speed of the number of new positive cases is almost the same if the vaccination effect is reflected. The timing of the start of decrease is different by two weeks to one month -> Reason: In machine learning, the applicability of the vaccination effect on the delta variant to Japanese people was not learned (limitation).
・Consideration of the decrease from August to September: (1) Speed of the effective reproduction number, which temporarily increased due to consecutive holidays, to return to the original level, (2) possibility that the decrease appeared to be drastic because of the synergy with the prevention effect of vaccination.
・Update of data and consideration (P6, 7) If flows of people decrease by 10% and 20% in Tokyo and Osaka, respectively, the effective reproduction number will be 1 or less (under the constraints: e.g. vaccine passport). The consideration of the impact of diffusion through a meal is not considered sufficiently in this method.