1
00:00:00,060 --> 00:00:07,500
Welcome everybody to the Department of Statistics. As Dorel just put it to me in this age when we don't need experts,

2
00:00:07,500 --> 00:00:18,030
actually the world is hanging on the voices of experts and we are lucky enough in this very worrying scenario to have two experts of our very own.

3
00:00:18,030 --> 00:00:23,550
So today, as many of you will know, two cases of the coronavirus have been confirmed in the UK,

4
00:00:23,550 --> 00:00:30,090
and our colleagues, Crystal and Robin, are going to give us some insights into world of real time epidemiology.

5
00:00:30,090 --> 00:00:34,380
This is what actually happens when you have thoroughly embedded in the subject.

6
00:00:34,380 --> 00:00:43,260
So I thought I would start by addressing what are the issues that we might want to know at this stage because this is a a new disease,

7
00:00:43,260 --> 00:00:47,010
a novel infectious agent. So what are the things you might start with?

8
00:00:47,010 --> 00:00:48,060
What are the symptoms?

9
00:00:48,060 --> 00:00:55,650
You have to design a case definition, you know, you're not going to test everybody that has every possible symptoms, so you've got to work on that,

10
00:00:55,650 --> 00:01:02,160
identifying what the causal agent is and Robin will speak about how you can compare the characteristics

11
00:01:02,160 --> 00:01:08,640
the symptoms with for this pathogen with what has been seen profiles and previous ones.

12
00:01:08,640 --> 00:01:12,330
How many cases are there? How many cases might there be in future?

13
00:01:12,330 --> 00:01:17,160
That is, of course, the key thing, but that's always projecting into the future is the what if.

14
00:01:17,160 --> 00:01:23,040
And while mathematics is very good at giving us the opportunity to think, what if under several different scenarios,

15
00:01:23,040 --> 00:01:28,980
because of course, we'll only just see one, there is inherent uncertainty and challenges in doing that.

16
00:01:28,980 --> 00:01:36,210
How serious might this disease be? So we could look at projecting future number of cases and thinking about how many people might ever be affected.

17
00:01:36,210 --> 00:01:40,440
But we are also interested in knowing what the case fatality ratio is.

18
00:01:40,440 --> 00:01:48,540
I mean, that's important in terms of measuring severity and determining what's a proportionate response, as well as how many people will be infected.

19
00:01:48,540 --> 00:01:53,310
But it also is important because people will start testing different treatments.

20
00:01:53,310 --> 00:01:58,730
And so knowing what the baseline case severity is is important.

21
00:01:58,730 --> 00:02:02,730
The incubation period distribution is important because if it turns out that I've

22
00:02:02,730 --> 00:02:07,380
exposed someone and I get find out tomorrow that I've got a particular disease,

23
00:02:07,380 --> 00:02:11,940
the incubation period is the time from when they were infected until when they show the signs of disease,

24
00:02:11,940 --> 00:02:15,360
and that knowing that distribution tells you if you're following them, how long.

25
00:02:15,360 --> 00:02:21,810
You should keep checking on Alison to see if she's still OK before you can give her the all clear.

26
00:02:21,810 --> 00:02:25,560
And finally, then of course, you're talking about the opportunities for control.

27
00:02:25,560 --> 00:02:37,820
We can talk characterise transmission, but what is the point of doing that if we don't also help identify and evaluate possible control measures?

28
00:02:37,820 --> 00:02:46,160
The numbers here are numbers that we made public through made with collaboration with colleagues at Imperial College London,

29
00:02:46,160 --> 00:02:50,330
where we have an infectious disease outbreak response unit.

30
00:02:50,330 --> 00:02:54,710
And so we collaborate with different people. So what we did here was follow up,

31
00:02:54,710 --> 00:03:02,120
similar work that we'd done originally when I was collaborating with Christophe and others on the influenza pandemic.

32
00:03:02,120 --> 00:03:08,360
So in that case, we were looking at the number of people who were coming out of Mexico and been confirmed with pendant.

33
00:03:08,360 --> 00:03:17,000
What then became pandemic influenza and comparing those to the number of people who had been identified as cases in Mexico.

34
00:03:17,000 --> 00:03:19,310
There was this mismatch.

35
00:03:19,310 --> 00:03:26,450
Of it seemed like more people were coming out infected than you might have expected, given the number of cases that have been reported.

36
00:03:26,450 --> 00:03:34,730
So we produced a report on the 17th of January where we looked at the fact that there had been three cases that had been exported.

37
00:03:34,730 --> 00:03:38,210
This is now in this case, exported out of China.

38
00:03:38,210 --> 00:03:48,760
And so in order to figure out what that means, so you can think of the population of infected people as being a sample of N or a population of size N,

39
00:03:48,760 --> 00:03:52,610
and then there's some possibility that the individuals who are infected,

40
00:03:52,610 --> 00:03:59,210
each of those individuals have a probability of travelling internationally and being detected as a case somewhere else.

41
00:03:59,210 --> 00:04:01,350
And so how do we think about doing that?

42
00:04:01,350 --> 00:04:09,320
One of the possibilities is we used data that we had previous access to on the number of people travelling internationally out of airports.

43
00:04:09,320 --> 00:04:14,720
And so you can look at Wuhan International Airport. You then have to think, OK, what is the number of people?

44
00:04:14,720 --> 00:04:17,330
What's the denominator for people who would use that airport?

45
00:04:17,330 --> 00:04:26,960
And we can consider either the the city population, which is 11 million or the metropolitan area population, which is 19 million.

46
00:04:26,960 --> 00:04:34,310
So this is a huge city and try to use that as the denominator for the probability that people would go out.

47
00:04:34,310 --> 00:04:40,190
Now there are all kinds of uncertainties here because we've assumed that everybody sort of the same in this.

48
00:04:40,190 --> 00:04:44,510
The probability that you got infected is is the same as everybody else.

49
00:04:44,510 --> 00:04:53,540
And the problem is so that the travellers are, assume a random sample of of the people there of that 19 or 11 million now.

50
00:04:53,540 --> 00:05:00,560
You know, all sorts of things about yourself that you're not you're probably not representative of everybody.

51
00:05:00,560 --> 00:05:05,390
And so, you know, people have raised the issue well for the market.

52
00:05:05,390 --> 00:05:11,690
Initially, those people who might have bought or gone to a market to look at wildlife, to eat,

53
00:05:11,690 --> 00:05:16,580
that they might be people who were more likely to travel if they were more affluent, for example.

54
00:05:16,580 --> 00:05:24,020
So there are uncertainties. But if we assume that everybody is equally likely and look at a detection window of the delay for cases,

55
00:05:24,020 --> 00:05:34,040
we estimated on the basis this baseline scenario, three exported cases implied central estimate of about 7500 cases that range between

56
00:05:34,040 --> 00:05:39,360
four hundred and four thousand when we updated the report on the 22nd of January.

57
00:05:39,360 --> 00:05:44,840
By that point, there were seven exported cases and that gave a central estimate of four thousand,

58
00:05:44,840 --> 00:05:52,820
and you see the sensitivity to the different assumptions based on the detection window and the catchment area.

59
00:05:52,820 --> 00:06:00,580
But to put this in context, the number of reported confirmed cases in China on the 16th of January was forty one.

60
00:06:00,580 --> 00:06:06,160
So that showed that they were would likely this analysis showed that there was an indication

61
00:06:06,160 --> 00:06:11,570
that it was very unlikely that there were only forty one confirmed cases at that point.

62
00:06:11,570 --> 00:06:20,170
Now that is not surprising because you would expect if it's a novel infectious agent, it takes time delays to identifying it, taking samples.

63
00:06:20,170 --> 00:06:24,100
In this case, sequencing it and identifying it as a new agent and so on.

64
00:06:24,100 --> 00:06:37,080
But just to give a sort of magnitude of what was going on. We subsequently looked at estimates where we looked at a range of are not value, so.

65
00:06:37,080 --> 00:06:42,620
We said, OK, so our not is the it's the basic reproduction number.

66
00:06:42,620 --> 00:06:48,620
It's the average number of people one infected person will cause in a fully susceptible population.

67
00:06:48,620 --> 00:06:55,740
If you think for this novel agent that the population is fully susceptible, it's being introduced.

68
00:06:55,740 --> 00:07:04,400
Question is what reproduction number would you have to have to get from a zoo, not a exposure, and you have to assume a particular date of that.

69
00:07:04,400 --> 00:07:08,150
Then you might say the end of December or some point in between there.

70
00:07:08,150 --> 00:07:14,000
What reproduction number would you have to have to be consistent with four thousand cases?

71
00:07:14,000 --> 00:07:20,360
And so the way we did that was through simulation and you see a set of simulations over there on the side.

72
00:07:20,360 --> 00:07:24,800
And that was for the case where they are not was equal to two point six,

73
00:07:24,800 --> 00:07:31,280
and that's where you get where the median of all those simulation gives four thousand cases.

74
00:07:31,280 --> 00:07:33,050
Now why do we look at four thousand?

75
00:07:33,050 --> 00:07:44,120
Well, that was because that was our central estimate that was published on the 22nd of January and corresponded to the 18th of January for the number.

76
00:07:44,120 --> 00:07:48,440
So what does this? What is two point six mean?

77
00:07:48,440 --> 00:07:54,380
Well, if you're going to control an infectious disease, you need to get the reproduction number down from whatever it was to start with.

78
00:07:54,380 --> 00:07:58,100
In this case, maybe two point six. You see different scenarios.

79
00:07:58,100 --> 00:08:06,560
The the lowest one comes when you assume that our lower limit, which is only one thousand cases,

80
00:08:06,560 --> 00:08:14,120
and that the initial number exposed at that market infected by the zoonotic source of the genetic sources,

81
00:08:14,120 --> 00:08:22,640
those animals infecting humans was two hundred. So to get from two hundred two thousand you don't need are not be that high.

82
00:08:22,640 --> 00:08:26,030
But if, for example, are not equals two,

83
00:08:26,030 --> 00:08:33,560
you need to prevent at least 50 percent of the transmission in order to control it and bring an increasing situation into a decreasing one.

84
00:08:33,560 --> 00:08:37,610
Obviously, all of us would like to bring the transmission to zero.

85
00:08:37,610 --> 00:08:44,260
And that's an hour of zero. But it doesn't have to get to zero in order to stop an epidemic.

86
00:08:44,260 --> 00:08:52,850
It just has to be pushed below one. Oh, I also wanted to talk about estimating the case fatality rate.

87
00:08:52,850 --> 00:08:57,080
So this is a figure that I pulled off of a Thai website.

88
00:08:57,080 --> 00:09:00,530
So, so the date of this was the 28th of January.

89
00:09:00,530 --> 00:09:05,930
And it was comparing the number of cases that have been seen so far for the confirmed for the coronavirus,

90
00:09:05,930 --> 00:09:09,860
which was just under three thousand and that there'd been eighty two deaths.

91
00:09:09,860 --> 00:09:14,750
So they took a ratio of those and said, OK, the mortality is two point eight percent.

92
00:09:14,750 --> 00:09:23,180
They compared that to sailors with the number of deaths and confirmed cases and said, OK, so ours is nine point six.

93
00:09:23,180 --> 00:09:28,790
So where did this come from? So if you look at worldwide, if you look at the W.H.O. website,

94
00:09:28,790 --> 00:09:38,030
you get these numbers off of for the eight thousand ninety eight cases and the number of deaths for sailors.

95
00:09:38,030 --> 00:09:44,900
I'll talk specifically about work that we did in looking at Hong Kong because that was Kristof and I were both

96
00:09:44,900 --> 00:09:50,000
involved in the analysis of Hong Kong data in collaboration with the University of Hong Kong colleagues.

97
00:09:50,000 --> 00:09:55,640
And there we found actually that the CFR was 17 percent, so considerably higher.

98
00:09:55,640 --> 00:10:00,660
So the rest of what I'm going to say, keep in that context.

99
00:10:00,660 --> 00:10:08,490
This shows you by day the number of cases in the Hong Kong SARS outbreak, so this is two thousand and three,

100
00:10:08,490 --> 00:10:19,020
so it was a an unusual, atypical pneumonia that had been detected in China in late 2002, almost exactly the same time of year.

101
00:10:19,020 --> 00:10:23,790
As things unfolded this year, this is how it unfolded in Hong Kong.

102
00:10:23,790 --> 00:10:31,290
So the first cases was the 22nd to February. Cases peaked in end of March, early April.

103
00:10:31,290 --> 00:10:37,290
So if you were looking at the number of deaths divided by the number of cases and you

104
00:10:37,290 --> 00:10:42,090
decided to do that on the 2nd of April because you look at that by the 2nd of April,

105
00:10:42,090 --> 00:10:50,200
there were nine hundred and twenty five cases. Almost eighty six percent of them were censored in that they were still in hospital

106
00:10:50,200 --> 00:10:54,980
and you didn't know if they were going to end up being a death or recovery.

107
00:10:54,980 --> 00:11:05,810
So there are then considerable risks associated with your estimate being the number of deaths so far divided by the number of cases so far.

108
00:11:05,810 --> 00:11:10,370
And that was how this was being tracked in terms of the case fatality ratio.

109
00:11:10,370 --> 00:11:14,630
So imagine that this said D divided by C.

110
00:11:14,630 --> 00:11:18,950
So that's method one is deaths so far divided by cases so far?

111
00:11:18,950 --> 00:11:22,430
We then considered alternatives. This is again collaborating.

112
00:11:22,430 --> 00:11:31,760
This is in two thousand three, and Method three was in collaboration with also with David Cox, who gave us, as always, very sage advice.

113
00:11:31,760 --> 00:11:32,870
But Method two.

114
00:11:32,870 --> 00:11:42,260
So one of the things that people said with such a simple estimate, it's so clear it's transparent cases divided by so deaths divided by cases.

115
00:11:42,260 --> 00:11:50,180
But if the time from becoming a case to dying takes a long time, there will be that gap where there's a delay and it will be biased downwards.

116
00:11:50,180 --> 00:11:55,310
An alternative that's almost as simple is deaths divided by deaths plus recoveries.

117
00:11:55,310 --> 00:12:05,930
And in that case, if the time from becoming a case to dying is similar in distribution to become a case to recovering, then this should be unbiased.

118
00:12:05,930 --> 00:12:09,260
We also looked, as I said, developing with David Cox,

119
00:12:09,260 --> 00:12:17,930
a method where we used its non-parametric and uses a Kaplan-Meier like estimate that is not that example is not four stars,

120
00:12:17,930 --> 00:12:24,080
but showing that you can look at the how the probability of recovery goes up and how the probability of survival goes down.

121
00:12:24,080 --> 00:12:29,390
And you end up somewhere in the middle with those two lines meeting when you've seen all of the outcomes.

122
00:12:29,390 --> 00:12:34,700
And so we looked at the performance of these looking back at the end of the SAR's epidemic,

123
00:12:34,700 --> 00:12:41,600
looking back so epidemic for Hong Kong and showed that if you look at the April second,

124
00:12:41,600 --> 00:12:46,010
so remember, that's when eighty six percent of the cases are censored.

125
00:12:46,010 --> 00:12:51,430
This is the estimate based on deaths so far divided by cases so far.

126
00:12:51,430 --> 00:13:00,100
And it has a very tight confidence interval. It's still a couple of percent this in black, this black square.

127
00:13:00,100 --> 00:13:08,770
This is the eventual case fatality rate, so between 13 14 percent for all the people that had become cases by the 2nd of April.

128
00:13:08,770 --> 00:13:16,310
And what you see if you just follow the one with the diamond, which is the simple one of deaths divided by cases over time, this will go up.

129
00:13:16,310 --> 00:13:23,780
It will eventually get to the right place if you observe all your outcomes because of all the deaths divided by all the cases.

130
00:13:23,780 --> 00:13:32,390
But it's very biased down initially our our simple alternative, which was death divided by death plus recovery.

131
00:13:32,390 --> 00:13:42,290
That's the one with the triangle here. And of course, it shows considerably more uncertainty because there weren't nearly 900 people who had outcomes.

132
00:13:42,290 --> 00:13:52,130
And so it's less certain, but it's not biased. If the time from becoming a case to dying was not similar to the time from being a case to recovering,

133
00:13:52,130 --> 00:13:58,190
which could happen if you were looking at having to hold people for a couple of weeks to be absolutely certain they weren't infectious,

134
00:13:58,190 --> 00:14:03,290
that can make a big difference. And then our fancy statistical one would be the way to go.

135
00:14:03,290 --> 00:14:09,260
How did that get interpreted? The press picked up that the estimate was increasing over time,

136
00:14:09,260 --> 00:14:14,870
that it was becoming deadlier if it was mutating to become more deadly, which is quite scary.

137
00:14:14,870 --> 00:14:22,850
And so we published our case fatality ratio estimates, and this is how it was reported on the 26th of April.

138
00:14:22,850 --> 00:14:30,260
So the research by a British scientist, it wasn't one, but there were several is due to be published and said that the virus could kill.

139
00:14:30,260 --> 00:14:31,820
So the case fatality ratio,

140
00:14:31,820 --> 00:14:40,870
the confidence interval at that point between eight and 15 percent W.H.O. had predicted or estimated between five and six percent at that point.

141
00:14:40,870 --> 00:14:47,650
And they said its estimate was more reliable because it was based on cases worldwide instead of just looking at Hong Kong.

142
00:14:47,650 --> 00:14:57,400
There was some discussion in the next hour by which the research was by Professor Roy Anderson, not a British scientist, but again and colleagues.

143
00:14:57,400 --> 00:15:04,450
But then this WTO spokesman said that Roy Anderson was a top class professional and his findings were probably accurate.

144
00:15:04,450 --> 00:15:11,620
So sometimes it's not the message. Sometimes it's the messenger.

145
00:15:11,620 --> 00:15:19,030
And I just wanted to show you before I hand over to Robyn. The case numbers as of today, it's nearly 10000 cases.

146
00:15:19,030 --> 00:15:23,260
This is by colleagues at Johns Hopkins.

147
00:15:23,260 --> 00:15:27,970
But I wanted to show you I have to try not to stand in the way.

148
00:15:27,970 --> 00:15:33,760
So if we look at deaths over deaths plus recoveries, it does not look good.

149
00:15:33,760 --> 00:15:38,320
But again, I hasten to add before people jump to conclusions on that.

150
00:15:38,320 --> 00:15:46,570
It's going to be very different if the time from being a case to dying is very different from the time, from being a case to recovering.

151
00:15:46,570 --> 00:15:51,550
And if we look at the examples, I look here because I can see them as opposed to looking at this angle.

152
00:15:51,550 --> 00:15:58,600
If you look at who by which is the first on each side, they've had two hundred deaths and one hundred and sixteen recoveries.

153
00:15:58,600 --> 00:16:04,300
But if you look at Guangdong. They've had zero deaths and 11 recovered,

154
00:16:04,300 --> 00:16:09,070
so it really depends on how long the epidemic has been there and also could vary depending

155
00:16:09,070 --> 00:16:14,650
on how long people are being who who aren't critical and dying are being kept in hospital.

156
00:16:14,650 --> 00:16:20,220
OK, I will hand over to Robyn this. Hi, everybody.

157
00:16:20,220 --> 00:16:27,000
Thank you very much for this opportunity to update you on the work that we've been doing in the Mathematical Institute to analyse this outbreak,

158
00:16:27,000 --> 00:16:32,310
as Crystal said, in real time. So Crystal put up this really great set of questions at the beginning,

159
00:16:32,310 --> 00:16:38,100
which are the kinds of questions that mathematical modelling can be used to answer in real time during an outbreak.

160
00:16:38,100 --> 00:16:42,480
I'm going to focus on two of those in terms of the work that we've done.

161
00:16:42,480 --> 00:16:46,950
The first one is this one here. So this was in the very first few days of the outbreak.

162
00:16:46,950 --> 00:16:51,840
And the question was, can we use modelling to identify what the underlying pathogen is?

163
00:16:51,840 --> 00:16:58,290
So can we say this is perhaps the smallest corona virus? Or is it instead something new that we've never seen before?

164
00:16:58,290 --> 00:16:59,910
So that was the first question.

165
00:16:59,910 --> 00:17:07,590
And then the second question that we've been focussing on more recently relates to this question of how many cases might there be and in particular,

166
00:17:07,590 --> 00:17:13,500
if the virus goes to a new location. So, for example, if it comes to the U.K., as it's now done,

167
00:17:13,500 --> 00:17:22,590
what's the chance of then seeing a chain of sustained human to human transmission in that new location as opposed to not seeing any more cases?

168
00:17:22,590 --> 00:17:26,880
So therefore that the two questions I'm going to kind of concentrate on.

169
00:17:26,880 --> 00:17:33,300
The first thing is, like I say, it was what we were trying to do in the very first few days of the outbreak and in particular,

170
00:17:33,300 --> 00:17:41,220
the first I heard of this outbreak was right back at the beginning of January. So I was alerted to it by colleagues at Hokkaido University in Japan.

171
00:17:41,220 --> 00:17:47,310
On what they'd seen is they'd seen this notice here in China and in particular, I don't speak Chinese,

172
00:17:47,310 --> 00:17:54,240
but I've been informed that this apparently says, well, there's been a cluster of atypical pneumonia cases in the city of Wuhan.

173
00:17:54,240 --> 00:18:00,060
But the source of these pneumonia cases, so exactly what the pathogen is, is as yet unknown.

174
00:18:00,060 --> 00:18:03,540
So I might, as I say, my Chinese isn't that good, but apparently that's what that says.

175
00:18:03,540 --> 00:18:07,590
And the other thing that was known at that time or what appeared to be the case

176
00:18:07,590 --> 00:18:12,030
was that all of those reported cases were linked to a particular seafood market.

177
00:18:12,030 --> 00:18:17,730
This one here the Huanan seafood market in Wuhan. So that was right back at the beginning.

178
00:18:17,730 --> 00:18:22,020
This is a kind of updates relating to the kinds of figures Krystal was just talking about.

179
00:18:22,020 --> 00:18:28,410
So back then, there were very few cases. I think something like twenty seven cases at the end of December, as you can see now,

180
00:18:28,410 --> 00:18:36,300
the number of reported cases has been accelerating quite rapidly. So going back, as I say, it's the beginning of the outbreak.

181
00:18:36,300 --> 00:18:41,850
This was the scenario, so cases kind of accumulated during December 2019,

182
00:18:41,850 --> 00:18:48,810
and when we got to the 31st of December, there were 27 cases of viral pneumonia that had been found.

183
00:18:48,810 --> 00:18:52,980
As I say, almost all of these cases were linked to this one on seafood market.

184
00:18:52,980 --> 00:18:58,320
And so with that in mind, there was no evidence at that point of any human to human transmission.

185
00:18:58,320 --> 00:19:04,590
So what was kind of suspected at that stage was that all of these people had picked up the virus from the market, from an animal reservoir.

186
00:19:04,590 --> 00:19:12,270
And that's as far as we knew. There was no evidence of any onward transmission between people.

187
00:19:12,270 --> 00:19:14,490
So what we could do at that stage, obviously,

188
00:19:14,490 --> 00:19:20,820
people want to do is they want to actually test the virus and find out exactly what the virus is that's driving the ongoing outbreak.

189
00:19:20,820 --> 00:19:25,710
And if you want to identify unambiguously what the virus is, that's what you have to do.

190
00:19:25,710 --> 00:19:27,490
Unfortunately, that takes some time.

191
00:19:27,490 --> 00:19:33,810
And so in just those first few days, we were saying, well, there's a huge amount of data that's available via social media sites,

192
00:19:33,810 --> 00:19:44,010
so via websites like flu tracker and other medical social media sites, and they give a rough idea of characteristics of the outbreak.

193
00:19:44,010 --> 00:19:48,840
As you can see, so some of some of the dates, you can see a kind of not necessarily that accurate.

194
00:19:48,840 --> 00:19:56,550
So for example, the fact that we thought maybe there wasn't any human to human transmission, but there was a vast quantity of possible data like this.

195
00:19:56,550 --> 00:20:03,030
So what I've got in this in this column here is a vector of what was observed at that point in the outbreak.

196
00:20:03,030 --> 00:20:08,640
So in particular. So there's a one indicating that there were lots of cases of atypical pneumonia.

197
00:20:08,640 --> 00:20:11,010
There's a zero here, which indicates that stage.

198
00:20:11,010 --> 00:20:18,600
It didn't appear like there was much human to human transmission and a whole bunch of other different characteristics that had been observed.

199
00:20:18,600 --> 00:20:21,990
So what you can then do is you can go back and look at previous outbreaks like, for example,

200
00:20:21,990 --> 00:20:27,870
the SAR's outbreak and you can write down an equivalent vector based on what was observed in that outbreak.

201
00:20:27,870 --> 00:20:33,840
And then what if what you want to do is you want to determine whether or not the current outbreak is indeed due to the source coronavirus?

202
00:20:33,840 --> 00:20:39,690
Essentially, what you want to do is a pairwise comparison between this vector here and this vector here.

203
00:20:39,690 --> 00:20:44,940
And if those two vectors are very similar, then you might conclude that in fact, the current outbreak is an outbreak of cells.

204
00:20:44,940 --> 00:20:50,640
If those two vectors are very different, then you might conclude that, well, perhaps it's not SAR's, and it's something a bit different,

205
00:20:50,640 --> 00:20:59,200
either one of the other possible pathogens that are on your list or in fact, not one of these and something that we've never seen before.

206
00:20:59,200 --> 00:21:04,750
So this is what we did and what it all essentially comes down to is quite a simple application of base rule.

207
00:21:04,750 --> 00:21:06,970
So the idea is what you want to say is you want to say, well,

208
00:21:06,970 --> 00:21:13,840
what is the probability that the outbreak you've got is due to some disease, for example, SA's given the observed characteristics.

209
00:21:13,840 --> 00:21:17,650
So given the vector of stuff that we've seen in the ongoing outbreak.

210
00:21:17,650 --> 00:21:20,080
And then if you apply based rule, you can say that, well,

211
00:21:20,080 --> 00:21:26,080
that's equal to the probability of seeing this set of observed characteristics, given that it is, for example,

212
00:21:26,080 --> 00:21:33,490
SA's times by an a priori probability that this outbreak is SA's divided by the equivalent terms for

213
00:21:33,490 --> 00:21:39,780
all of the possible candidate diseases on your list of diseases that it could hypothetically be.

214
00:21:39,780 --> 00:21:43,440
So anyway, we kind of did this calculation, you have to make a few assumptions, in particular,

215
00:21:43,440 --> 00:21:48,810
you have to make some assumptions about the distance between these vectors and how that relates to the

216
00:21:48,810 --> 00:21:53,010
probability that the observe characteristics are due to the disease that you're comparing against.

217
00:21:53,010 --> 00:21:56,880
So I would say that at the moment, this is quite a rough kind of method.

218
00:21:56,880 --> 00:22:04,020
But what this rough method gives you is it gives you a kind of priority list of potential pathogens that could be causing the ongoing outbreak.

219
00:22:04,020 --> 00:22:09,150
So unless that looks a little bit like this going from the one that appears to be most likely given the characteristics

220
00:22:09,150 --> 00:22:15,500
you've observed two diseases that are in fact less likely pathogens that are less likely to be the causal agent.

221
00:22:15,500 --> 00:22:21,050
I mean, if you want to see what you can do is you can extend this analysis and you can say, well, what given these vectors,

222
00:22:21,050 --> 00:22:28,910
what is the probability that in fact this outbreak is being driven by a pathogen that is not any one of those pathogens that are on your list?

223
00:22:28,910 --> 00:22:29,600
And if you do that,

224
00:22:29,600 --> 00:22:37,580
then what you obtain is you obtain a kind of risk score for the probability that this that this outbreak is driven by none of these pathogens.

225
00:22:37,580 --> 00:22:41,930
And in particular, what you can see here is this is called disease X. This is how I refer to this.

226
00:22:41,930 --> 00:22:48,530
This means a pathogen that you haven't previously seen what you see as you see something that changes your probability of disease.

227
00:22:48,530 --> 00:22:52,760
X is something that changes throughout the outbreak as more information comes in.

228
00:22:52,760 --> 00:22:58,160
So as you see more characteristics, that suggests that this perhaps isn't stars, for example,

229
00:22:58,160 --> 00:23:04,600
then you'll probability that this is in fact an entirely new thing is something that jumps up through time.

230
00:23:04,600 --> 00:23:10,300
OK, so that was the first kind of bit of analysis that we did, the second one is something that we've done more recently.

231
00:23:10,300 --> 00:23:14,770
And so like I said, the question here is, well, if in the U.K. you go somewhere else.

232
00:23:14,770 --> 00:23:23,100
So if a case goes to a new location, what's the probability that we then see sustained transmission in that new location?

233
00:23:23,100 --> 00:23:27,360
And so I should say this figure has nothing to do with coronavirus at all, but I just wanted to demonstrate the idea.

234
00:23:27,360 --> 00:23:35,270
So this is a model that many of you might be familiar with. So this is a simulation of an FBI all model and in particular, a stochastic RSI all model.

235
00:23:35,270 --> 00:23:40,560
The idea here is that you flip a coin lots of times and according to the result of each coin flip,

236
00:23:40,560 --> 00:23:44,460
you either generate a new infection in your population or you assume that one

237
00:23:44,460 --> 00:23:51,030
of the infected individuals is that you've got recovers or becomes isolated. In either case gets removed from the current outbreak.

238
00:23:51,030 --> 00:23:52,500
And so you flip a coin lots of times.

239
00:23:52,500 --> 00:24:00,590
And if you do this once, according to a very simple model like this, you might see an outbreak projection that looks a little bit like this one here.

240
00:24:00,590 --> 00:24:07,310
If you do exactly the same thing again, under absolutely identical conditions, you might see something that looks very slightly different.

241
00:24:07,310 --> 00:24:14,800
So like this blue line here, this blue line is exactly the same thing, but very slightly different sequence of coin flips.

242
00:24:14,800 --> 00:24:21,250
But then if you do this again. So once again, on the absolutely identical conditions, then you might instead see something a bit like this.

243
00:24:21,250 --> 00:24:23,980
So this is starting with one infected case and in the red case,

244
00:24:23,980 --> 00:24:29,530
it just happens to fade out and you don't see an outbreak driven by sustained human to human transmission.

245
00:24:29,530 --> 00:24:32,290
And again, this is just based on the coin flips that you have.

246
00:24:32,290 --> 00:24:37,210
So if your first coin flip is someone getting removed rather than someone generating a new infection,

247
00:24:37,210 --> 00:24:40,250
then you'd see a very small outbreak like the red one here.

248
00:24:40,250 --> 00:24:46,210
So this leads kind of naturally, this is just to introduce the idea that when someone brings the virus to a new location,

249
00:24:46,210 --> 00:24:51,430
there's some probability that what you see is an outbreak driven by sustained human to human transmission.

250
00:24:51,430 --> 00:24:57,010
So in other words, you see something like the blue or the black, as opposed to something that looks like the red.

251
00:24:57,010 --> 00:24:58,930
And so if you had a forecasting model in theory,

252
00:24:58,930 --> 00:25:04,630
what you could do is you could simulate this lots and lots of times count the proportion of times you get something like the blue and the black,

253
00:25:04,630 --> 00:25:07,180
as opposed to the proportion of times you get something like the red.

254
00:25:07,180 --> 00:25:13,640
And that gives you then the probability of getting sustained transmission when the virus is in a new location.

255
00:25:13,640 --> 00:25:21,340
That's kind of the idea. It turns out of simple models like the iPhone model, you can calculate this probability analytically.

256
00:25:21,340 --> 00:25:24,340
So in fact, for the asylum model, it comes down to solving a quadratic equation.

257
00:25:24,340 --> 00:25:30,400
But for more complex models, it comes down to doing more complex calculations. But essentially what you see is something like this.

258
00:25:30,400 --> 00:25:36,340
The probability of sustained transmission is some function of how transmissible the pathogen driving the outbreak is.

259
00:25:36,340 --> 00:25:41,680
And if you have a very transmissible pathogen, then the probability of getting a sustained outbreak is quite high.

260
00:25:41,680 --> 00:25:48,520
Whereas if you have a pathogen that's not very transmissible, then the probability of getting a sustained outbreak is substantially lower.

261
00:25:48,520 --> 00:25:53,890
And this measure of pathogen transmissibility on the x axis, so Crystal referred to this earlier,

262
00:25:53,890 --> 00:25:58,450
but this is just the product of the rate at which new infections are being generated by

263
00:25:58,450 --> 00:26:03,760
each infected hosts and the length of time that an infected host is infectious for.

264
00:26:03,760 --> 00:26:09,790
So in other words, if you have if you have a pathogen for which hosts are infectious for a very long time,

265
00:26:09,790 --> 00:26:12,040
then obviously the pathogen is going to be very transmissible.

266
00:26:12,040 --> 00:26:17,380
Similarly, if you have a pathogen for which infected hosts generate new infections very quickly,

267
00:26:17,380 --> 00:26:21,640
then again you get a pathogen that's very transmissible. So that's kind of the idea.

268
00:26:21,640 --> 00:26:25,570
We applied some of these methods to data from the ongoing outbreak.

269
00:26:25,570 --> 00:26:30,190
So in particular, one of the things that we wanted to quantify was how long individuals are infectious.

270
00:26:30,190 --> 00:26:36,880
Four. And so the way that we did that was we took data from forty seven patients from very early in the outbreak.

271
00:26:36,880 --> 00:26:45,340
The data that were recorded were the number of days in between individuals showing symptoms and then being taken to hospital and isolated.

272
00:26:45,340 --> 00:26:51,490
So in other words, this is the measure of how long people could have been out in the community and spreading the virus for.

273
00:26:51,490 --> 00:26:55,300
So this gives you some sort of metric of of the duration of infection. And then from that,

274
00:26:55,300 --> 00:27:03,610
you can use methods like the ones I just showed you to calculate the probability of getting a sustained outbreak when the virus goes to a new place.

275
00:27:03,610 --> 00:27:07,960
And then what we did is we looked at, well, what happens if you take this distribution here and you squash it?

276
00:27:07,960 --> 00:27:14,290
So in particular, what happens if you can cut down the amount of time between symptom onset and isolation?

277
00:27:14,290 --> 00:27:20,500
So, for example, by having more rigorous surveillance of symptomatic hosts and if you have more rigorous surveillance of symptomatic hosts,

278
00:27:20,500 --> 00:27:25,550
then your risk of getting sustained outbreaks is substantially smaller.

279
00:27:25,550 --> 00:27:33,470
And so, as I said today, this is kind of quite a topical analysis because today we've obviously seen these two cases in the UK.

280
00:27:33,470 --> 00:27:38,750
This is one of the first preprints, one of the first analyses of these outbreaks that was put out online and so slightly.

281
00:27:38,750 --> 00:27:41,300
Strangely, this is my first experience of any press attention.

282
00:27:41,300 --> 00:27:47,720
So in particular, this graph appeared on BBC News Live from my living room, which is kind of strange.

283
00:27:47,720 --> 00:27:55,280
But anyway, so the idea is that now this is a very topical analysis. And so I guess these kinds of calculations are particularly relevant right now.

284
00:27:55,280 --> 00:28:01,200
So just one thing to say then is about an extension to this work that we're now looking at.

285
00:28:01,200 --> 00:28:05,150
So my PhD student, Francesca Reid, is now doing some work on this in particular.

286
00:28:05,150 --> 00:28:06,800
That's in response to this article here.

287
00:28:06,800 --> 00:28:15,830
And evidence that came from China to what's been found now is that some infections appear to be happening before the in fact becomes symptomatic.

288
00:28:15,830 --> 00:28:23,570
So there some possibility that infections can spread, not from people that are showing signs of having the disease, but instead before that time.

289
00:28:23,570 --> 00:28:32,160
And with that in mind, this distribution here isn't a very good characterisation, perhaps of of how long people are infectious for.

290
00:28:32,160 --> 00:28:37,190
So what we're doing is we're updating this analysis so that we don't have individuals that have the disease

291
00:28:37,190 --> 00:28:42,920
going from a state in which they're not symptomatic and not infectious to being symptomatic and infectious.

292
00:28:42,920 --> 00:28:45,050
So this was our assumption before.

293
00:28:45,050 --> 00:28:51,890
But now we've got something slightly different in which individuals go from being not symptomatic and not infectious to being not symptomatic,

294
00:28:51,890 --> 00:28:57,680
but potentially infectious. And then it's only later that they become both symptomatic and infectious.

295
00:28:57,680 --> 00:29:01,100
And so you can then extend the kinds of analysis that I just showed you.

296
00:29:01,100 --> 00:29:07,760
The key conclusion that came out of this is the exactly how you might look to try and control

297
00:29:07,760 --> 00:29:13,310
pathogen with different levels of pre-symptomatic transmission might be very different.

298
00:29:13,310 --> 00:29:19,370
So if you have a pathogen that for which a huge proportion of infections are occurring before symptoms occur,

299
00:29:19,370 --> 00:29:21,410
then that might that might affect what you do.

300
00:29:21,410 --> 00:29:28,460
In particular, a control strategy that shortens the period of symptomatic infectiousness might not there might not be very effective any more.

301
00:29:28,460 --> 00:29:28,850
So actually,

302
00:29:28,850 --> 00:29:36,890
just going and trying to find symptomatic people might not be sufficient if there's a lot of pre-symptomatic infection to try and control the virus.

303
00:29:36,890 --> 00:29:42,610
So I'm going to leave it there. I'm going to Haslam's Christophe and I look forward questions at the end of things.

304
00:29:42,610 --> 00:29:49,830
So I think I'll always remember the day when Crystal gave a couple of us who

305
00:29:49,830 --> 00:29:55,840
were working with her on such a quick lesson in survival analysis and bias,

306
00:29:55,840 --> 00:30:02,470
and right truncated data showing that the CFR estimates case fatality rates estimates one,

307
00:30:02,470 --> 00:30:07,870
two and a half percent, but between seven and 15 percent where she didn't say was at the time.

308
00:30:07,870 --> 00:30:19,300
There was also a change in the case definition for the inclusion in databases, which doubled the amount of reported cases that afternoon.

309
00:30:19,300 --> 00:30:24,580
So it was a memorable afternoon and illustrates the two points,

310
00:30:24,580 --> 00:30:34,030
which is that important decisions need to be made for controlling an epidemic with very imperfect observations.

311
00:30:34,030 --> 00:30:41,740
And governments need to expose themselves to making difficult decisions where it's very

312
00:30:41,740 --> 00:30:47,140
hard to say whether they were the right decisions or not on the basis of incomplete data.

313
00:30:47,140 --> 00:30:52,600
So the ability to try to adjust and to come up with statistical estimates,

314
00:30:52,600 --> 00:31:01,180
which are corrected for bias is absolutely critical for improving the nature of decision making.

315
00:31:01,180 --> 00:31:11,680
So one of the sort of, having worked out mentioned just very briefly two perspectives one from the SA's epidemic and one from the 2009

316
00:31:11,680 --> 00:31:20,470
flu pandemic in terms of some of the really difficult questions that need to be answered with imperfect data.

317
00:31:20,470 --> 00:31:30,820
So one of the things I think that's that's new this time is a lot of prior knowledge and very open and rapid sharing,

318
00:31:30,820 --> 00:31:35,860
both of data of analysis and of genome data.

319
00:31:35,860 --> 00:31:45,160
So the first thing is, I think there's an interesting question as to whether we should be sort of thinking about this in a slightly Bayesian manner.

320
00:31:45,160 --> 00:31:51,910
There are sort of priors that can help us sort through the fog of uncertainty and the genetic data from

321
00:31:51,910 --> 00:31:58,960
this new virus was shared within days and showed two things to sort of data points which really help.

322
00:31:58,960 --> 00:32:08,020
So the first is that the genetic data from the first sequences from Wuhan and from travellers were basically all genetically identical,

323
00:32:08,020 --> 00:32:16,600
separated at most by a single mutation and a relatively large genome in error prone and replication.

324
00:32:16,600 --> 00:32:28,030
So we're pretty confident that the virus hadn't been circulating cryptically for months and really did emerge at the earliest and late November,

325
00:32:28,030 --> 00:32:37,360
and probably most likely in mid-December. So that's different from previous outbreaks, where the cryptic transmission,

326
00:32:37,360 --> 00:32:45,130
unreported transmission both in the case of SaaS and in the case of the flu pandemic had been going on for weeks or months.

327
00:32:45,130 --> 00:32:51,310
So this really is sort of Real-Time reporting at the same time in the epidemic for those previous epidemics.

328
00:32:51,310 --> 00:32:56,680
None of these analysis had even started, and none of the data had been shared.

329
00:32:56,680 --> 00:33:05,020
So to put that in context. And the second thing is that the origins of both this new coronavirus are not supposed to stand,

330
00:33:05,020 --> 00:33:11,170
and the new corona virus and of the SARS virus are both within the same clade of bat viruses.

331
00:33:11,170 --> 00:33:19,540
So they share a common origin. And the genetic distances between the new coronaviruses and the source viruses are relatively small.

332
00:33:19,540 --> 00:33:25,660
They're small enough that they're less than the difference between influenza A and influenza B.

333
00:33:25,660 --> 00:33:31,030
And we would both consider those to be the same virus in the case of HIV.

334
00:33:31,030 --> 00:33:37,360
So taxonomic taxonomy, there was a legitimate question as to whether there should be Soulsby or a new corona virus,

335
00:33:37,360 --> 00:33:45,910
and they've settled on new coronavirus. But a prior might have been to assume or might be to assume that this is SA's like.

336
00:33:45,910 --> 00:33:50,560
Now what's the big difference between SA's and influenza that's relevant?

337
00:33:50,560 --> 00:33:57,310
First of all, with SA's containment was a sort of massive success story.

338
00:33:57,310 --> 00:34:07,870
And serological studies after the outbreak confirmed that there was basically no asymptomatic transmission and very little transmission in children.

339
00:34:07,870 --> 00:34:13,070
So the isolation measures and the quarantine measures and the contact tracing measures,

340
00:34:13,070 --> 00:34:18,580
we were relying on transmission that occurred before symptoms on the x axis.

341
00:34:18,580 --> 00:34:26,800
Here I've got the proportion of infections that occur or transmissions that occur prior to symptoms or due to asymptomatic infection.

342
00:34:26,800 --> 00:34:32,530
And on the y axis I've got are not. So you'll forgive the graphics, which were from 2004.

343
00:34:32,530 --> 00:34:41,890
I hesitated about whether to update them, but basically SaaS was fairly infectious but fairly easy to control because most transmissions occur.

344
00:34:41,890 --> 00:34:48,580
After symptoms, as Robin said, so being able to estimate that a clear indicator of that is that lots of

345
00:34:48,580 --> 00:34:53,110
transmissions occurred in hospitals and health care workers where increased risk.

346
00:34:53,110 --> 00:34:59,770
So the influenza estimate here is completely wrong. From this paper in 2004, the transmissibility is much lower.

347
00:34:59,770 --> 00:35:08,620
But there's transmission from very mild cases. So the challenge in 2009 was that it was clear that transmission was going on.

348
00:35:08,620 --> 00:35:14,380
It was clear that there were severe cases, but there was no idea when the critical decision needed to be made,

349
00:35:14,380 --> 00:35:18,850
whether the severe cases were one in 100 cases, one in a thousand cases.

350
00:35:18,850 --> 00:35:27,460
As it turned out, the estimate was probably one in 100000. Infections was leading to a severe to a severe case.

351
00:35:27,460 --> 00:35:35,200
And therefore, in the case of influenza, the decision was gradually made to actually not try to control the virus because more harm was

352
00:35:35,200 --> 00:35:44,260
being done by the social isolation measures needed to control the virus than the virus itself.

353
00:35:44,260 --> 00:35:54,790
And there's some juicy mathematics that you can engage with to to predict the impact of isolation and contact tracing

354
00:35:54,790 --> 00:36:01,120
showing what what the main of this parameter space leads to infection control for different measures or not.

355
00:36:01,120 --> 00:36:06,940
So this is important for deciding whether you're going to go for a reactive control

356
00:36:06,940 --> 00:36:12,350
policy where you follow up symptomatic cases or whether you need to go for a mass.

357
00:36:12,350 --> 00:36:23,080
As as Robyn mentioned, a mass control policy where you just reduce contacts between all individuals in order to contain the epidemic.

358
00:36:23,080 --> 00:36:28,000
So the question is, is this new Kovarik coronavirus, which is genetically similar to us?

359
00:36:28,000 --> 00:36:36,040
Is it a virus that has very high pathogenicity with very few mild infections, very few pre-symptomatic infections?

360
00:36:36,040 --> 00:36:37,780
Or is it to the right of here?

361
00:36:37,780 --> 00:36:48,040
Or is it all of the way like HIV, where most transmissions occur from pre-symptomatic or mildly symptomatic individuals?

362
00:36:48,040 --> 00:36:53,950
This very impressive New England Journal of Medicine paper was published yesterday afternoon,

363
00:36:53,950 --> 00:37:01,600
which is an analysis of the first 450 or so cases from where your hand might be starting to give some indication.

364
00:37:01,600 --> 00:37:08,020
And I refer you to the paper loss estimates of our note, which are consistent with we have published.

365
00:37:08,020 --> 00:37:11,890
But one critical observation here is that in the latest data,

366
00:37:11,890 --> 00:37:21,250
nearly three quarters of all people who have been reported as infections and were hand have not did not at the point of being interviewed,

367
00:37:21,250 --> 00:37:28,300
have exposure to either the market where the infection started or a person with respiratory symptoms.

368
00:37:28,300 --> 00:37:34,090
So there is these are the first data set to move us away from the source prior

369
00:37:34,090 --> 00:37:38,710
to a more important role for pre-symptomatic and asymptomatic transmission.

370
00:37:38,710 --> 00:37:42,400
There's also been well documented cases, such as the cases in Germany,

371
00:37:42,400 --> 00:37:51,220
where an asymptomatic individual who then develops symptoms travelling back to China had infected two people within in Germany.

372
00:37:51,220 --> 00:37:57,340
So the reality is that we still have considerable uncertainty about the severity of infection and the severity.

373
00:37:57,340 --> 00:38:04,000
Infection could be close to the case fatality rate, which is clearly quite high as as Crystal has been indicated,

374
00:38:04,000 --> 00:38:10,930
or it could be much lower if there are many more mild or asymptomatic infections.

375
00:38:10,930 --> 00:38:16,240
So one of the lessons of 2009 is it took nearly four months to sort out this uncertainty,

376
00:38:16,240 --> 00:38:22,930
during which time many decisions had to be made and some statistical lessons were very useful.

377
00:38:22,930 --> 00:38:29,530
And in particular, these are the the labels here refer to the flu to 2009 flu,

378
00:38:29,530 --> 00:38:36,460
but this was a statistical paper by statisticians and was revisited by several different groups

379
00:38:36,460 --> 00:38:41,830
and showing the property of estimates so-called sort of hierarchical pyramidal estimates,

380
00:38:41,830 --> 00:38:47,110
and showing that you can converge on robust estimates very quickly by multiplying the

381
00:38:47,110 --> 00:38:52,930
numbers much more quickly than by doing very large surveys and well-designed surveys,

382
00:38:52,930 --> 00:38:57,190
looking at sort of the degree of asymptomatic infection and then looking at the

383
00:38:57,190 --> 00:39:02,170
degree of severe infection and so on and multiplying those numbers together.

384
00:39:02,170 --> 00:39:08,000
So I think right now, in order to make the key decisions, in addition to the data showing,

385
00:39:08,000 --> 00:39:16,060
we sort of very urgently need to know these kind of data which are going to be reliant on screening the appropriate populations at risk.

386
00:39:16,060 --> 00:39:20,041
I think you.