Projected early spread of COVID-19 in Africa

For African countries currently reporting COVID-19 cases, we estimate when they will report more than 1 000 and 10 000 cases. Assuming current trends, more than 80% are likely to exceed 1 000 cases by the end of April 2020, with most exceeding 10 000 a few weeks later.

The World Health Organization (WHO) declared COVID-19 a public health emergency of 23 international concern (1) and then a pandemic (2), citing its rapid global spread and risk of 24 overwhelming healthcare services with patients requiring critical care. As of 24 March 2020, 25 WHO situation reports (SITREPs), indicated 45 African countries reported at least one 26 laboratory-confirmed infection ("reported case") of COVID--19 (World Health Organization 27 2020). Reported cases underestimate actual infections due to the mix of mild symptoms (3, 4), 28 the similarity of symptoms common to the region (5), and weak surveillance (6). However, 29 assuming constant reporting activity, reported cases grow in proportion to the underlying 30 epidemic, and even with under-ascertainment of the number of actual cases, reported cases 31 provide a useful indicator of stress on healthcare systems. We can use this surrogate for the real 32 epidemic to forecast future trends, and understand the consequences of a slow public health 33 response and what preparations need to be made now. 34 . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https: //doi.org/10.1101//doi.org/10. /2020 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript.

The Study 35
We use a branching process model to project the number of future cases of COVID-19 in each 36 country. This model assumes each case produces a number of new cases (distributed 37 . We start with cases corresponding to the first 25 39 (or fewer) reported cases in the WHO SITREPs up to 23 March 2020 (10). Using those epidemic 40 parameters and initial cases and dates, we simulate the accumulation of the reported cases. We 41 assume there are always sufficient unreported infections to continue transmission, and that new 42 cases represent a reporting sample from both identified and unidentified transmission chains. As 43 long as a constant reporting fraction persists during this period, and unreported spread is large 44 relative to reported cases (or reporting has negligible impact on control), this is a reasonable 45 approximation. 46 For each set of country-specific initial conditions, we generate n=10 000 epidemics, discarding 47 any that fade out, consistent with our assumption of unreported transmission chains. We identify 48 the dates when each simulation run crosses 1 000 and 10 000 reported cases, and then evaluate 49 the 50% and 95% quantiles of those dates to determine the forecast interval. The model was built 50 in the R statistical programming language, using the bpmodels package (11), and the 51 data2019nCoV package for the SITREP data (12). All analysis code is available from 52 https://github. com/SACEMA/COVID10k. 53 . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript.
We project that almost all African countries are likely to pass 1 000 reported cases by the end of 54 April 2020 and 10 000 within another few weeks ( Figure 1 and Table 1); alarmingly, these are 55 largely synchronized continent-wide and real disease burden will certainly exceed reported 56 cases. Since our projections assume failed containment of initial cases and no interventions 57 reducing early transmission, they are pessimistic relative to any benefits of local action. 58 However, containment measures, e.g. travel restrictions, increased testing, contact tracing, 59 isolation of cases and quarantine of contacts, are likely to slow, but not halt, real epidemic 60 growth (13). Indeed, increased testing may accelerate the time to reporting these numbers, as 61 improved ascertainment increases the identified fraction of real cases. However, the model also 62 optimistically assumes surveillance capacity is not overwhelmed or stymied, which would slow 63 reaching 1 000 reported cases while the real disease burden grows uncontrolled. Because we 64 ignore these effects, the model is only appropriate for short-range forecasts. 65 As model validation, we applied this same forecasting approach to countries world-wide that 66 have now exceeded 1 000 reported cases; we did not consider those with more than 10 000 cases, 67 as they have all undergone substantial control measures modifying epidemic growth. We found 68 that 44% of actual reporting dates fell within the 50% prediction intervals, and 79% within the 69 95% interval (Figure 2), indicating the forecast prediction interval is too certain, as expected for 70 a rapid but low detail model. We further showed that forecast performance is not a random 71 outcome by performing a randomization test: we shuffle the assignment of forecast days-to-1 72 . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript. 000-cases to different countries, and score 1 000 shuffled predictions; the real forecast score is 73 significantly different from random at the p < 0.001 level (Figure 2 inset). 74 Specific to Africa, the forecast for South Africa fell within the 50% prediction interval (SITREP 75 69; 29 March 2020). From 23 March 2020, we projected a few other countries would also likely 76 be crossing this threshold soon: Algeria, Egypt, Morocco, Senegal and Tunisia. As of SITREP 75 77 (4 April 2020), the first three are still fast approaching this limit, while fast and intense responses 78 in the latter two may have successfully slowed the epidemic. 79

Conclusions 80
Using reporting to date, and assuming similar epidemiological trends to those seen globally, we 81 project that almost all African countries are likely to exceed 1 000 reported cases by the end of 82 April 2020, and 10 000 within another few weeks. This timing is largely synchronized continent-83 wide and real disease burden will certainly exceed reported cases. Our projections assume no 84 substantive changes between the initially reported cases and the forecast points; while some 85 countries have taken drastic actions, many have not or have acted slowly. As seen in other 86 regions, because onset of severe symptoms can be delayed weeks from infection and last several 87 weeks, interventions have limited immediate impact on new hospitalizations or facility demand, 88 meaning that most of the countries in our projection would be well past 1 000 cases by the time 89 the effects of interventions started today would be observed. 90 . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi. org/10.1101org/10. /2020 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript. . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript. . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10.1101/2020.04.05.20054403 doi: medRxiv preprint . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101 Reserved space. Do not place any text in this section. Include the mandatory author checklist or your manuscript will be returned. Use continuous line numbering in your manuscript.

Uganda
Apr 15 indicates randomization results for actual forecast (red) versus randomly assigned forecast 128 (grey), with 0.975 quantile indicated by the line. For actual reporting dates, 44% fell within the 129 50% prediction intervals, and 79% within the 95%. 130 . CC-BY-NC-ND 4.0 International license It is made available under a author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
is the (which was not peer-reviewed) The copyright holder for this preprint . https://doi.org/10. 1101