Projects on the topic of COVID-19 hospitalisations - mostly DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
The primary analysis presented in this report is a single page the presents the actual data on hospitalisation, ICU, ventilation and deaths, together with a simple "forecast". This exploits the observable relationship between each of these events, that typically occur for a given percentage of cases, after a typical/average delay. The exact % and delay are subjective - I've set my best guesses as the defaults, but you can try alternatives.
For deaths, the cumulative total of cases is used. For the other series, the total of the last 14 days' cases are used. This follows the differences in how those figures are reported by health departments - deaths are a cumulative total whereas hospitalisation, ICU and ventilation cases are daily snapshots. Previously the "active cases" reported by each health department was used, but those were subject to arbitrary classification differences.
A separate page is presented for each outbreak, by Geography (State/Territory) and start month. It's interesting to compare differences in the % Expected between Geographies and outbreak phases.
Inspiration for this analysis came from this Twitter post, referencing work by Nigel Marks. Thanks also to this Twitter post by Richard from Sydney which pointed out that deaths should be derived from cumulative/total cases.
Interactive dataviz - adjust the parameters to suit.
Time-shifted Distribution analysis Link to interactive DataViz
Hospitalisation Link to interactive DataViz
ICU Link to interactive DataViz
Choose any state/territory, date range etc. Adjust the parameters to suit, default is 18% after 6 days.
I now consider the remaining pages to be less useful, so I dont intend to develop them further. Automated data updates will continue.
I noticed a few international analyses using a similar method - comparing:
- "expected hospitalisations" - % X of new cases (typically 2% to 10%), vs
- hospitalisations, shifted back N days (typically 2 to 10 days)
The expectation is that the lines track each other, confirming the hypothesis that % X of cases result in hospitalisation after N days.
Needing to using different % X and N days values might reflect changes in vaccination rates, or the characteristics of variants.
Here's one by Oliver Johnson for the UK, using 5% and 10 days:
Another by Andrew Steele for the UK, using 8% and 10 days:
The lines tracked quite tightly until vaccination rates increased, then hospitalisations started to drop away.
And another by Eran Segal for Israel, using 2.2% and 4 days:
The lines tracked quite tightly for their first and second waves (top chart), but the bottom chart shows an even stronger effect of vaccinations.
The idea being explored is that cases can predict future need for hospitalisations, ICU, ventilation and death. Parameters are used for:
- the % of cases predicted to need hospitalisation, ICU, ventilation and result in death
- the number of days to shift/delay between case reporting and hospitalisation, ICU, ventilation or death
Once parameters are found that get a close fit between the Actual and Forecast series, then any divergence might indicate something has changed in the situation, e.g.
- new variant causing more/less severe illness
- new variant causing faster/slower onset of severe illness
- health system under pressure
- more/less vulnerable population becoming infected
Australian COVID-19 data from covidlive.com.au.
A page is presented set for the current New South Wales and Victoria outbreaks, with an indicator of the date that outbreak started.
An "explorer" page is also presented where you can choose to focus on a state of interest or alter the date range.
I found the closest fit for the line seems to be for 10% (cases expected to result in hospitalisation) and 6 days (delay from case reporting to hospitalisation). Using the interactive dataviz you can adjust those to suit your understanding.
For ICU, I found the closest fit for the line seems to be for 5% (cases expected to require ICU) and 10 days (delay from case reporting to ICU admission). Using the interactive dataviz you can adjust those to suit your understanding.
Note the hospitalisation figures used internationally are understood to be the number of people admitted to hospital each day. That figure is publicly available in some countries, but not for Australia (AFAIK - please contact me if you are aware of a source dataset).
Instead in Australia we can get the number of people currently in hospital/ICU each day. So for a fairly close approximation of hospital admissions, I am using the daily increase in hospitalisations. This will be slightly understated at the tail end of each outbreak, when more people are leaving hospital than entering. I add the daily deaths figure, assuming that those were previously reported as hospitalised and in ICU. Australian data is sparse and spiky so I smoothed using 7-day rolling averages.
Note Queensland current policy is to admit all COVID-19 cases to hospital, so their hospitalisation figures dont make sense for this analysis.
THIS REPORT IS NOT HEALTH ADVICE - REFER TO YOUR LOCAL HEALTH AUTHORITY.