Thread by @ojwatson92

Our report investigating the under-reporting of COVID-19 deaths in Damascus, Syria was released this morning. This analysis has changed how I view the extent to which COVID-19 has potentially spread unobserved in many parts of the world. 1/n

https://twitter.com/MRC_Outbreak/status/1305801730177851392

Firstly, 2 main takeaways:

1. Best estimate is that 1 in 80 deaths due to COVID-19 have been reported as of 2nd September, suggesting that as many as 4,380 deaths may have been missed.
2. Epidemic in Damascus significantly more advanced than reported deaths would suggest 2/n

But how do we arrive at these estimates, in a setting that has been ravaged by war for nearly a decade leading to both weakened surveillance and health systems?

We start with excess mortality data. 3/n

On the 2nd August, the Damascus Governorate released a statement (facebook.com/damascusgovrno…) in response to social media reports of large increases in recorded mortuary cases. The statement reported the number of all cause deaths for the region between 25th July - 1st August 4/n

During this 8-day period, 832 deaths were reported giving an average of 104 deaths per day.

This is significantly in excess of estimates of expected daily mortality, which by the same Governorate office have suggested is between 15-50 deaths per day. 5/n

Collating multiple consistent reports of baseline mortality, we computed approximate excess mortality, which we used to infer the level of under-ascertainment of COVID-19 deaths in Damascus. 6/n

We use a transmission model framework to fit to these reported daily deaths (below), BUT with the assumption that reported deaths represent a proportion of true total COVID-19 deaths. We then infer the proportion of deaths that are reported by comparing to the excess deaths. 7/n

Scanning across a range of under-reporting levels, we find that assuming between 1% - 3% of deaths explains the excess mortality shown here (points and triangles showing two assumed baseline mortalities, small vertical bars showing reported daily deaths). 8/n

Given the extensive changes to Syrian demography and health services resulting from war, it is hard to confident of a number of model parameters (number of beds, population size and demographics, fatality ratios etc). 9/n

In response, we explored 1320 parameter sets. Summarising across this sensitivity analysis, we find that 1.25% of deaths are reported, with a likely sensitivity range between 1% - 3%: 10/n

To explore the reliability of the model fits, we looked at reports of when hospitals were reported to be at capacity in Damascus. Reports vary, but between 17th July - 30th July. We find that 1% and 1.25% under-reporting results in correct timing of capacity being reached. 11/n:

Using the best estimate of under-reporting (1.25% deaths reported) produces a starkly different picture of the epidemic in Damascus as of 2nd September, With an attack rate (total proportion of population infected) equal to 40%. 12/n

HOWEVER - it is important to be aware of the numerous difficulties in trying to characterise the epidemic in Damascus and the assumptions we have made and the related caveats. The main 3 of these are: 13/n

1. Daily reported deaths follow the same shape as the epidemic after accounting for under-reporting.

2. Under-reporting is fixed over time.

3. Deaths that occur within hospitals and outside hospitals are equally likely to be reported.

14/n

For example, let's alter the 3rd assumption and say that only individuals who die in hospitals may be captured by reported deaths. If we do this we produce a very different epidemic, which explains all 3 sources of data (reported deaths, excess deaths, hospital capacity). 15/n

Here we show the earlier fit the three sources of data in gold, and our new approach in blue. This assumes that when hospital capacity is reached, reported deaths will flatten because deaths are only reported from within hospitals: 16/n

In response, we went looking for an alternative data source that could give us another view of the epidemic and help decided between these different epidemics trajectories.

For this, Mervat Alhaffer pointed me in the direction of a Facebook group - facebook.com/wafiatdimashq 17/n

Traditionally, when individuals die in Damascus, a paper certificate of death is printed and affixed to
household walls in nearby neighbourhoods of the deceased, with details of the deceased’s relatives and the consolation events being held. 18/n

In response Mervat helped translate a number of certificates to estimate:

1. How many are from people who have died in Damascus (sometimes certificates are made for people who died outside Damascus in rural areas)

2. The delay from death to certificate upload. 19/n

@ZackAkil wrote code that helped detect which images were certificates. He did this using Google Cloud’s Vision API (@Googlevision -cloud.google.com/vision). What seemed like an impossibly long manual task, took a few hours and resulted in identifying 18,038 certificates. 20/n

Despite the informality of the certificate mortality, mortality trends in this alternative data source reveal a consistent baseline of 300 certificates per month in 2017 - 2019. In contrast, 809 and 1066 were uploaded in Jul and Aug 2020, with clear excess mortality 21/n

We repeated, our model fitting to this data, which confirmed our initial picture of the epidemic in Damascus and our main conclusions:

1. A small fraction of COVID-19 deaths have been ascertained and reported in Damascus.
2. The epidemic is substantially more evolved

22/n

This study provides evidence that COVID-19 epidemics may have occurred largely unobserved in many places. In response, work needs to be done leveraging alternative data sources if we hope to better understand how widespread the pandemic may be. 23/n

This work would not have been possible without substantial help from from many individuals. In particular, Mervat Alhaffar (@LSHTM), Zaki Mehchy (@SyRG_SCaHR), @ZackAkil, Francesco Checchi (@LSHTM) and the COVID-19 group @MRC_Outbreak 24/n

Lastly, big thank you to @ejbeals for starting this whole study and her work covering Damascus and Syria. Most importantly, thank you to the anonymous Syrian doctors, health officials, epidemiologists, academics who were instrumental and wish to remain anonymous for security.

@threadreaderapp unroll