For the past few years, the strains selected for the flu vaccine turn out to be not the ones prevalent. Why do we have such low predictability on which strain will be the most infectious and can AI improve the predictability ? Or will this always remain a shot in the dark ?
The first thought is to increase the number of strains in the vaccine.
The second is to vastly increase production capacity so as to shorten the time needed for production, so better information is available at this later time.
And third, following up the second, to give a second flu shot later in the year encompassing these later strains
.
How do you think AI could help? Any kind of learning heuristic needs a lot more cases than we get with the flu. There is also not a stable population to learn from, since mutated flu viruses will sometimes not have been seen before.
Production capacity isn’t the problem, it is producing the right vaccine. Look how long it took with Covid despite an unprecedented effort. If you start when you identify this year’s flu strain for sure, you are not going to have a vaccine until the season is over. That’s why they have to guess.
It seems fairly logical to me that the strains of flu we vaccinate against would be become less prevalent than other strains, it’s possible that the strains we vaccinate against would have been the most prevalent if we had not vaccinated against them.
The problem with the flu is that we are trying to predict a moving target. Influenza mutates all the time. So there are always new strains appearing, and at the time the vaccines begin production we need to lock in which of the known strains to pick to vaccinate against, knowing that by the time the vaccines are ready the world will have moved on. New strains may have arisen, and the strains in existence at the time the vaccine was designed may not be most dominant ones come next winter. About the best we can do is have look at the other hemisphere - ie in our summer we look at those countries that are in winter, and check which strains look nasty and pick from them the ones we will put in the vaccine. Hard to imagine AI of any kind doing any better than random dumb luck. This simply isn’t the sort of problem any sort of AI is good at.
The flu strains compete with one another to some extent, plus the weather, patterns of travel, and new mutations, all make for a chaotic spread of the various strains. You use the best knowledge you have, and make your best guess. A bit like betting on the ponies. Success is not dissimilar.
I’m sure this is a big part of it.
Another big part is probably the high mobility of the population relative to what was possible even 30 years ago, let alone back in 1945, before the highway system, when the yearly flu vaccine first became a thing. Yes, I’m aware that they existed before that, but the wide-scale distribution, and the PSAs promoting them for everyone began after WWII.
Knowing for sure which strains will be most prevalent would require knowing where the reservoirs of disease are that are going to spread to everyone else. But if we knew where those reservoirs were, we could completely eradicate them, through a combination of much smaller-scale vaccination and selective quarantining.
Granted that flu viruses mutate; but bear in mind that a vaccine against X can still be effective against X(mutated) - instance the current discussion about COVID mutations and how there is reasonable confidence that this won’t seriously impact vaccine efficancy.
So, with that in mind, here’s how flu vaccines are actually arrived at. For vaccines for the northern hemisphere, the latest Flu season in the southern hemisphere is tracked to see what viral strains are prevalent. Based on this information, in the early part of the year the WHO will make a recommendation for strains to be used in vaccines for the upcoming northern hemisphere flu season - like this:
https://www.who.int/influenza/vaccines/virus/recommendations/2020-21_north/en/
-so that this year’s recommendations include stuff like this example (egg based, quadravalent):
Egg-based Vaccines
- an A/Guangdong-Maonan/SWL1536/2019 (H1N1)pdm09-like virus;
- an A/Hong Kong/2671/2019 (H3N2)-like virus;
- a B/Washington/02/2019 (B/Victoria lineage)-like virus; and
- a B/Phuket/3073/2013 (B/Yamagata lineage)-like virus.
This is where my knowledge starts to thin out a bit. I have always assumed that you can go out and buy an [attenuated] A/Hong Kong/2671/2019 (H3N2)-like virus off the shelf, and this is going to provide reasonable protection against that strain of flu and recent mutations thereof. So to make a flu vaccine. a manufacturer will have their basic flu vaccine vehicle, for which they will buy the strains the WHO recommends, drop them into the vehicle, go through an expedited registration process for this year’s flu vaccine - and you have your vaccine.
The bottlenecks are supply chain - someone has to manufacture those off-the-shelf antigens that every manufacturer will be chasing; and then it takes time to manufacture the season’s supply of vaccine. So for a winter flu season the WHO is recommending strains nearly a year in advance (28 Feb 2020 in this case). That’s a long way out to be making a prediction about the strains which will be prevalent for the next season.
So: as I understand it, the time required to produce a vaccine means that the WHO has to make a recommendation early (start of southern hemisphere winter), which makes their predictions less accurate.
j