Aims
The RESPINOW-Hub is a collaborative
project on nowcasting and short-term forecasting of the incidence of
respiratory diseases and the resulting healthcare burden in Germany. The
aims of the project are twofold:
- Real-time operation: Firstly, we aim to provide a resource
to support the assessment of current and expected future epidemic trends
in real time.
- Methods development: Secondly, we aim to develop, assess
and compare statistical methods for these purposes.
The project covers data on multiple
respiratory diseases (invasive pneumococcal disease,
RSV, seasonal influenza) as well as combined indicators from syndromic surveillance.
Currently, the project is in a beta phase, where much of the
computational infrastructure is operational, but prediction models are
still in development.
While this platform does not address SARS-Cov2/COVID-19, it is
inspired by projects on COVID-19, e.g., the German COVID-19
Hospitalization Nowcast Hub and the European COVID-19 Forecast
Hub.
The data
In the following, the data sources underlying the
RESPINOW-Hub are summarized.
Arbeitsgemeinschaft Influenza
The Arbeitsgemeinschaft
Influenza sentinel surveillance system
consists of more than 600 general practitioners, who on a voluntary
basis provide information on the number of consultations for respiratory
infections. Reporting is done directly to RKI either electronically (SEED-ARE)
or by fax. We use the consultation incidence for acute respiratory infections (ARI; ICD-10 codes J00
– J22, B34.9 and J44.0) per 100,000 inhabitants. This indicator is not
specific to one pathogen and thus forms part of syndromic surveillance. Data are in principle
available per age group and region (with certain pairs of German states
merged), but we currently only use data at the national level.
Further information:
Clinical Virology Network
tba
Further information:
ICOSARI
Around 70-80 hospitals from a large hospital operator (Helios
Kliniken GmbH) report new hospitalizations due to severe acute respiratory infections (SARI). This
hospital sentinel system covers 13 out of 16
German states and 5-6% of all hospitalizations in Germany. SARI is
defined according to a set of ICD-10 diagnostic codes (J09 – J22),
meaning that the indicator is syndromic not
specific to one pathogen.
Further information:
SurvStat
For a large number of communicable diseases, laboratory-confirmed
cases are notifiable in Germany. Local health authorities
(Gesundheitsämter) receive reports from general practitioners and
laboratories, and via state-level administrations (Landesbehöden)
forward them to Robert Koch-Institute. The SurvStat system aggregates the
resulting data. In our platform, data on invasive
pneumococcal disease, RSV (only covering the state of Saxony) and
seasonal influenza are displayed. Data are available by state and
age group (note that we aggregated certain pairs of states so regions
agee with the regions of Arbeitsgemeinschaft Influenza).
It should be noted that laboratory analyses are on only performed for
a fraction of all patients, meaning that SurvStat only covers
part of the actual disease incidence. Reporting completeness can vary
over time and depends on many aspects, including healthcare seeking
behaviour and reimbursement policies.
Further information:
Virological Surveillance (National Reference Center)
Roughly 20% of the sentinel GP practices participating in
Arbeitsgemeinschaft Influenza are equipped to perform nose swabs. They
collect samples according to a set of symptoms and an age
stratification. Samples are tested for a variety of pathogens, but we
only display data on RSV and seasonal influenza.
Further information:
The models
Most models are still under development and do not feed into the
operational platform yet. Currently, the following models are displayed
(in alphabetical order):
KIT-epinowcast
This nowcasting model combines a latent random walk model for the actual
epidemic curve with a parametric delay model. Inference is done in a
Bayesian fashion. The implementation is based on the R package epinowcast and can be found
here.
KIT-hhh4
This is a simple seasonal count time series model as implemented in the
R packages surveillance
and hhh4addon. It is
built on top of the KIT-simple_nowcast and currently only applied to
SARI data (influenza having too irregular seasonality to be captured
well by this model). The implementation is available here.
KIT-simple_nowcast
This nowcasting model is based on simple multiplication factors based on
recently observed reporting delays. Uncertainty intervals are based on
past nowcast errors using a parametric negative binomial model. A model
description (for a slightly more complex setting of daily data) is
available in the supplement of Wolffram et al (2023). The
implementation can be found here.
RIVM-KEW
This model is a simplified version of the model presented by van de Kassteele,
Eilers and Wallinga (2019). The reported counts by date and delay
are described by a negative binomial distribution. The expected values
are modelled by a two-dimensional P-spline surface and other covariates.
This surface is extrapolated for all dates and delays outside the
reporting triangle. Model fitting is done using the mgcv
package in R.
Glossary
In the following we provide brief explanations of relevant terms. A
similar list (in German) is available on the website of
Arbeitsgemeinschaft Influenza.
Acute respiratory infection (ARI)
Acute respiratory infection (ARI) is a summary term for illness from
various respiratory pathogens and refers to a combination of symptoms
(“acute pharyngitis, bronchitis or pneumonia with or without fever”).
The exact definition is based on ICD-10 diagnostic codes (J00 – J22,
B34.9 and J44.0; see e.g., here, page 86; in
German). More information can also be found here
on the website of Robert-Koch-Institut.
Collaborative forecasting
Forecasting infectious disease spread and the resulting healthcare
burden is challenging and experience shows that different models may
lead to rather different forecasts. It is thus considered good practice
to use several independently operated forecasting models in parallel;
see e.g., this editorial by Reich
et al (2023). This enables the identification of reliable models and
the combination of different forecasts into so-called ensembles
predicitons (see below).
Ensemble prediction
Ensemble predicitons are combinations of forecasts from different models
or methods. These have been found to be more more reliable than
individual models in many disease forecasting efforts (see e.g., Cramer et al ,
2022). They are also very commonly used in other domains like meteorological
and economic forecasting.
Nowcasting
Most epidemiological indicators are subject to reporting delays, meaning
that events which happened in a given week only appear in the respective
data set some time later. This means that the most recent data points
are often incomplete and will still be subject to upwards corrections.
This can lead to an artificial dip and the wrong impression of a
downward trend. Statistical nowcasting serves to anticipate thiese
corrections and thus determine relevant trends in real time. A good
overview on the topic is given in the paper by Günther
et al (2020).
Invasive pneumococcal disease
Pneumococcal
diseases are caused by the bacterium Streptococcus pneumoniae. While
a large part of the population is colonized by Streptococcus
pneumoniae, i.e., carries the bacterium in their body, this does usually
not lead to illness. Invasive pneumococcal disease (IPD) refers to
symptomatic cases where the bacterium can be isolated from normally
sterile sites.
Respiratory Syncytial Virus (RSV)
Respiratory
syncytial virus (RSV) is a common respiratory disease which
typically causes mild forms of infection, but can be more severe for
young children and elderly persons. Vaccines against RSV have become
available in recent years.
Sentinel surveillance
Sentinel
surveillance (see also here
in German) refers to a surveillance scheme where selected general
practitioners, hopsitals etc. provide more detailed information on the
occurrence of infectious diseases than required in the mandatory
reporting schemes. Based on the catchment areas of these healthcare
providers, estimates for disease activity or healthcare burden in the
entire population can be obtained.
Severe Acute Respiratory Infection (SARI)
Severe acture respiratory infections represent a subset of respiratory
infections with particularly strong symptoms. The exaxt definition is
again formulated in terms of ICD-10 diagnostic codes (J09 – J22; see
e.g., here,
page 86; in German). Hospitalizations due to SARI are subject to
sentinel surveillance in Germany, see above and e.g., here.
Short-term forecast
Forecasts of infectious disease spread are typically only feasible for
rather short time horizons (see e.g., here for a
detailed account). Depending on the characteristics of a disease and the
epidmeiological situation this duration typically reaches from a few
days to several weeks. Based on experience, we consider 2-3 weeks a
reasonable maximum forecast horizon for the respiratory diseases in
question. Beyond these horizons, so-called scenario projections are used
to make statements under various sets of assumptions (see e.g., Howerton et al
2023).
Syndromic surveillance
Syndromic
surveillance is a summary term for surveillance techniques that are
based on symptoms of affected persons rather than the identification of
the causative pathogen. The surveillance of ARI and SARI are examples of
syndromic surveillance.
Getting involved
The RESPINOW-Hub is open to new collaborators. If you are
interested in the modelling of respiratory diseases (whether based in
Germany or abroad), do not hesitate to get in touch.
RESPINOW-Hub is part of the RESPINOW Consortium, which in turn is part of the MONID network, funded by the German Ministry of Education and Research.