Symptom Watch UK

Counting Methodology

The Twitter Stream.
We received English language tweets that were filtered for covid19 symptoms using MedDRA terms for fever, cough and anosmia. Explicit covid19 terms such as coronavirus were not included in the search list.
We remove official retweets or tweets that start with #RT.

Selection of UK tweets.
Tweets that we determine have originated in the UK are selected via a series of tests.
Some tweets are geotagged. This provides us with the location of the tweeter.
For others we look at the profile of the author to ascertain whether they have provided an appropriate location.
Finally we run a trained classifier on the tweet text to differentiate between those texts that are likely to have been written by someone in the UK and texts that are not. (Note this test does not provide a location within the UK and so is not represented in the figures on this site.)

Are the tweets reporting symptoms?
The selected UK tweets are then put through a classifier that determines whether the tweeter is reporting their own current symptoms (class 1), reporting their own past symptoms (class 2), reporting someone else’s current symptoms (class 3), reporting someone else’s past symptoms (class 4), or are not reporting symptoms at all (class 0).
If the tweet is not reporting symptoms it is rejected. Presently the sum of classes 1-4 are shown in the tables.

The mapping of tweets to areas.
Tweets are mapped from their original location, (city, town, village), to lower tier local authorities, upper tier local authorities, NHS regions and countries using lookup tables provided by The Office for National Statistics.

Server downtime.
For every fifteen minute period we check the number of tweets received. If zero then the server is deemed to be down for that period. The number of tweets reported = count * 96/(96 - downtime periods) and the date is annotated as date¹. On the rare occasion when the server is down all day, (May 17th 2020) then zero counts are reported and the date is annotated as date².

Back to top.