The index below is a live metric of “Hospital Stress” that represents: “how much hospitalization numbers are growing daily in percentage”. By hospitalization number i mean, hospitalized_symptomatic_cases and intensive_care_cases (see section below for explanation) Equally important, this index is ALSO a decent(*) indicator, but lagged(5to8days), of the real hidden recent growth of the ACTUAL contagion.
This index is also a proxy of the “percentage growth” of the contagion, we speak about the growth of contagion NOT the absolute number. As of today the real number of infected is highly uncertain (possibly 5 to 20 times bigger than the official numbers of swab/tested positive officially comunicated by the Protezioe Civile). Nonetheless, even with the Proxy, for today decision making the daily percentage change of all these numbers is very useful.
Governments, Society as a whole and individuals are trying to push this indicator it down to zero, and then below zero. It is possible that its value will linger around zero for few days that means the new cases entering hospitals are balancing the one exiting (including exiting for home-care, not fully recovered). The Trendline are a visual help to predict where we would be is the dynamic of the system would be equal. We should not expect to be equal due to many non-stationary exogenous effects and just border-effects that may happen even after containment. That said its deceleration (angle) has been fairly stable.
I will update this numbers and figures with daily data. The recent trend has slightly slowed down in the last week but improved on the latest days. The peak for hospial stress has happened around the 5th Apr (possibly 2 days before). Note that I clean/smooth out some of the single day high-frequency oscillations noise, but the indicator is still as fast as possible to react.
In figure 2 this index and its deceleration is plotted disaggregated for the North, Centre and South regions.
If a second wave of infection may arise, this index will restart to grow, also if its value “sits around” a small but positive number it is an informative sign that the SARS-Cov2 could become endemic. As of today we have some control to try to avoid that.
(*) Finally, it is important to keep in mind that IF hospitals saturate their space (with death numbers still growing) the index also would abruptly spike to zero with just two days lag (this may have happened in certain areas of the north, Bergamo for example, but I have not analysed that local data). One would need to check local version of this index. A small numbers of local saturation should not ruin the index information too much.
UPDATE: 07/08-April Data, and COMMENTS:
As of the 22nd March the prediction of a peak for Hospital stress was around the end of March. We probably had that peak more like between 2nd and 4th of April. The Deceleration intensity has been less strong near the end of March. You can see a bit of pick up in hospitalization in the south Region in the next Figure below.
For the peak of REAL hidden New daily contagions, the number should have peaked approximately 5 days +/- 3 before, around the 26th of March. Note that the descend from the peak of contagion bell shaped curve (not plotting this here, but a zero in growth of contagion is the sign of the arrival on the peak), is expected to be slower than the ascent. So it could take another 30 days to have near zero NEW contagion cases.
The deceleration was fairly constant during the period of March. From 33% to 1%, from 1st March to 1st of April, decelerating by 1% per day approximately. When Reaching the Zero Percent growth the social-distancing measures seem to have less effect. Will we have a slower deceleration (that means the hospitals will get empty). It seems yes. It seems natural since now the number of people PRESENT in hospitals and Intensive care is dominated more by the speed/time to recovery (or to die), than the number of NEW patients/cases entering the hospital structures. The government should collect the numbers of NEW cases entering a state (such as NEW hospitalized, NEW Intensive care) to have proper indicator. I write more about it in the section in the comments.
GUESSTIMATING REAL CONTAGION NUMBERS: We can estimate the real hidden number of contagion, using any assumption/model on the real number for the 1st of March, THEN we can compound its growth with this indicator in time: cumprod(1+G(t)), where G(t) is starting from 33% on the 1st of MArch . With my data the TOTAL growth of REAL contagion cases , between the 1st of March and the 1st of April is in the range [ 60 to 100 ] times depending on how much noise we want to put. ( I could/should do a Montecarlo MCMC but I have no time now, it would just be an “elegant dress” for the assumption on the initial distribution of the error anyway). If you say real infected on the 1st of march were 5 times the total_hospitalized on the 8th of march (between 3000 and 5000). Hence using midpoints we can say 5* 4000 = 20000, growing 80 times, then you get Between 1.6 Millions infected as of the end of March. It would be around 3 Millions with the highest range of the estimates.
As mentioned above, the value of the decreasing blue line/area in that Figure1, is a measure of daily percentage growth of what we can call the “hospitals stress” ADDITIONALLY it is also a (lagged) indicator/signal of the recent growth of the REAL contagion.
Society and governments are trying to push it down to zero and below (it started around 35%, since the 6th of Apr is below 0%, and decreasing only very slowly since). Its value is an ensemble average of daily-percentage-change of 3 times series, like a combination of 3 signals: hospitalized_symptomatic_cases, intensive_care_cases, and their sum. To understand beter Fig1 consider the following time series related to the COVID-19 epidemic currently ongoing in Italy:
– A) Daily-percentage-change-of-hospitalized_symptomatic_cases
– B) Daily-percentage-change-of-intensive_care_cases
– C) Daily-percentage-change-of-total_hospitalizations
There are two evident reasons to use percentage change: 1) this series tend to grow exponentially, at least locally, that is new daily cases are proportional to number of cases from the previous day; 2) secondly this hospitalizations are in general a percentage of the actual unknown total infected cases (appearing in the data, with a lag after incubation and worsening conditions). It is difficult to backtrack the number of daily contagion (even if it could be done better with better data see comment section below). However even without knowing the real number of contagion we can estimate its daily percentage change, using as a proxy the percentage change of the various hospitalizations.
To compute it, first I pre-smooth the A), B), C) times series cutting higher frequencies noise with a robust smoothing.
Then average their smoothed value.
Finally to highlight the trend, as a benchmark, I plot a couple of linear extrapolation, using a robust-weighted-rolling linear-fit, to estimate when we can expect to have a zero or negative growth/change, hence when the “hospital stress” would stop growing, that’s when the hospitalization would peak (and with some lag also the new infected).
With data as of the 27th Mar, we could “hope” the stress to stop growing between the 29th March and 1st of Apr. However UPDATE, we see that the deceleration slowed: and with This is NOT a good sign it increase the risk that the infection will become endemic slowly going at .
There are many models for epidemic dynamics as well as ways to fit the wave of new contagion/hospitalizations (using functional forms such as Gaussians, logistic, Gompertz, or other non linearities), hence multiple ways to estimate a future date for the peak. However this indicator does the same type of prediction as well as being useful to estimate the dynamic of the stress also if/when second “waves” or new hotspots of infections may arise (that would be more complicated with fixed curve modelling). The extrapolation being linear is also quicker to grasp visually.
The indicator can also help in managing an exit strategy from the current lockdown. Helping to assess as quickly as possible, the effectiveness of policies to re-start social and economic life. In Particular as a fast metric working with multi waves scenarios.
The metric is also robust to single (or a couple) data point spikes/errors/delays due to all sorts of operational reasons such as reporting or operational capacity saturation.
In the Figure 2 I compare the same indicators but divided for the North, Centre, South macro regions.
Note that even if the growth stays at 5% level, at the mid stage of epidemic, it would be a highly undesirable scenario, we need the line to go negative as fast as possible to start to have a reduction in hospitalizations. Also, in case there is a full saturation of the hospital systems, the indicator may hit/stay abruptly the zero line and the dead number (with a delay) would keep growing.
The epidemic started in the Lombardia region, so the grouping below is both geographical as well as approximately represents the epidemic at different stages (even if in Piedmont or Liguria there was approx a week delay of the onset). In terms of policy the starting times for social distancing were similar (8th / 9th Mar). This region definitions maybe overlapping. Here in Fig2, I am plotting the actual real dates with no sinthetic lags.
North_ish= Lombardia,Piemonte,Veneto,Valle d’Aosta,P.A. Bolzano,P.A. Trento,Friuli Venezia Giulia,Liguria, Emilia Romagna
Center_ish=Toscana,Umbria,Marche,Lazio,Abruzzo,Molise
South_ish=Basilicata,Campania,Calabria,Molise,Puglia,Sardegna,Sicilia
The data in the south_ish area is more noisy and sparse, the growing rate smoothed value recently does appear to start slowing down more significantly. I also Put the indicator for All_wihtout_Lombardia and Lombardia-alone. In Lombardia some province ( Bergamo, Brescia, some part of Milano) have had frequent saturations based on other info. However the informativeness of the index as a whole at the Italian Level stll holds, as it is seen from the center and south trends or Italy without Lombardia. It basically starts at around 30/35 % and ends aroudn 2 % between the beginnnig and the end of March. Approx 1% per day average deceleration of percentage growth. This should hold for the contagion itself.
(If you know the Italian data already you can skip the next section. You can click on the little arrows to expand certain sections)
Click for Expanding The Data Source Section
Raw Data from “Protezione Civile Italiana” (better sources are welcome if you have them)
https://raw.githubusercontent.com/pcm-dpc/COVID-19/master/dati-regioni/dpc-covid19-ita-regioni.csv
Here there is a list of daily pdf reports.
https://github.com/pcm-dpc/COVID-19/tree/master/schede-riepilogative/regioni
The time series with daily frequenceies are : “hospitalized_symptomatic”,“intensive_care”,“tot_hospitalized”,“home_isolation”,“tot_positive”,“recovered”,“dead”,“tot_cases”,"“tested”
in Italian (“ricoverati_con_sintomi”,“terapia_intensiva”,“totale_ospedalizzati”,“isolamento_domiciliare”,
“totale_attualmente_positivi”,“dimessi_guariti”,“deceduti”,“totale_casi”
“tamponi”)
| hospitalized_symptomatic | intensive_care | tot_hospitalized | home_isolation | tot_positive_now | new_positive | recovered | dead | tot_cases | |
|---|---|---|---|---|---|---|---|---|---|
| 2020-04-17 | 25786 | 2812 | 28598 | 78364 | 106962 | 3493 | 42727 | 22745 | 172434 |
| 2020-04-18 | 25007 | 2733 | 27740 | 80031 | 107771 | 3491 | 44927 | 23227 | 175925 |
| 2020-04-19 | 25033 | 2635 | 27668 | 80589 | 108257 | 3047 | 47055 | 23660 | 178972 |
Click for Expanding this Section
There are various models for epidemic data (the SIR with its varations being the most used, see all the references in [1], [16]) here we just want to build a live indicator (Blue line in Fig1 or Fig2) to monitor the current state of the epidemic without assuming any model or stationarity in the time series dynamic, or for their relationship and parameters.
The Figure3 below is just the raw data, and shows visually that in the logarithmic space there is an “approximately” linear growth of the time series. It also appears the linear growths, the tangent lines of this plots, tends to flattens out with time. Independently from stationarity assumptions, one can measure the daily percentage growth of the time series, such as an “instantaneous” growth rate that (hopefully) decreases with time, due to all sorts of effects, both intrinsic of the epidemic and exogenous such as government policies, as well as other noisy contributions.
NOTE: The raw numbers in Figure 3 represent the current number of people/cases in a specific “state” (e.g. a “state” is “being hospitalized”). In Figure 4 the daily-differences of the Log(Series) of Figure 3. It would be better to have daily number of people entering each state, such as daily number of new_hospitalized, or daily number of new_intensive_care. One person/case can, in principle, move from hospitalized to intensive_care and back, between hospitalized and non-hospitalized, as well as from intensive care to dead (usually 50% of the cases according to some research [see summary reference [1]]). Also, some recovered from icu and clinically recovered are still contagious as well and is not so clear where they have been classified. The number of New tested positive and new death cases are in principle computable, since those are “one-way” state, however they are: 1)total positive cases: have upper bounds dependent on will/ability of large scale testing. 2)dead: very slow and small proportion, with a large dispersion of lag: death has a big lag-uncertainty in the from symptoms (median 20 +/- 10 days ). Also some cases may go straight to the dead state never being recorded as hospitalized or in intensive care. New Intensive care could be partially deduced from the change in total current intensive care numbers adding the new_dead (I may do this later), but other uncertainties holds and this flows between states have different timescales (icu to dead seems to be median of 5 days, while recovering from icu may take weeks).
All data dynamics, in general, can be influenced by “border effects” like saturation of available space/resources to deal with fast local growth of the cases. Moreover there are occasional and regional dependent delays in noting the data points and the inability to even record/attribute the cases for death/hospitalization etc. With anecdotal knowledge some data points maybe cleaned ad-hoc (not yet done here). A lot of issues may be present in the data, such as large underestimation for positive cases and even related deaths (see [17])
We can estimate the time-trend of growth doing the 1-day differences of the numbers of each Series above : 100*(Log(Series(t))-Log(Series(t-1))) . this is simple the 1-day percent Change. Plotted in the next figure Figure4. There are two evident reasons to use percentage change: 1) this series tend to grow exponentially, at least locally, that is new daily cases are proportional to number of cases from the previous day; 2) secondly this hospitalizations are in general a percentage of the actual unknown total infected cases (appearing in the data, with a lag after incubation and worsening conditions). It is difficult to backtrack the number of daily contagion (even if it could be done with better data see comment sections).
However even without knowing the real number of contagion we can estimate its daily percentage change, using as a proxy the percentage change of the various hospitalizations.
See also the Global Trend Line that highlight the tendency of all series. Indeed percentage changes allows to put all Series in teh same natural scale. They could be compared/averaged.
Comments on data collection strategies.
This data collected by the Italian “Protezione Civile” is useful, but far from ideal. Hospitals are overwhelmed with work and sometimes they lag with the publications of data. Would be useful to have hospitalization data from other countries, and for comparisons, and at that stage it would be important to have harmonization of definitions between countries, for all type of cases. Definitely at least for the definition of death associated to Covid-19. Currently Germany death numbers are defined with different criteria that in Italy. Transparent Numbers and criteria for all testing including tested negative, types of intensive care (invasive, non invasive), etc.
One of the most important live analysis for decision making would be to build backward the time series of daily New-Infected cases. This is currently not-explicitly observable, but with better data can be reconstructed with a certain lag (incubation period to symptoms is 6 +/- 3 days)[3] [7]. Using all the measured series Positive cases, …until dead Again as mentioned above, also for that type of backtracking it would significantly help if all hospitals keep/disclose not only the current total numbers of cases in a specific “state”, but more importantly the new-arrival for each state. For obvious reasons this is computable for death, and tests cases, but its unclear for different level of hospitalization. Hospitals/Government should ask to patients, collect and release :
When was each case having the first symptoms ? [7].
New entry in each level of hospitalization ( And possibly increase the details of the level of hospitalization: early-pneumonia, non-invasive int-care, invasive-int-care, etc.).
In general record any New entry in any “state” that may be two-way state.
In addition for any good analysis on positive cases, countries need to do more testing (as in South Korea) and in particuar randomized population testing, To assess any risk factor, for infectivness, temperature, humuidity, social context, as well as Health- Risk-factors. This is a key part of the Test-Trace-Contain procedure that has been successful to contain the growth of the infection in many Asian countries.