Lower Respiratory Infections: GBD 2019

Disease Description

Lower respiratory infections (LRI), principally clinician-diagnosed pneumonia and bronchiolitis, is a major global killer of both children and adults. Symptoms include shortness of breath, weakness, fever, coughing and fatigue. It is important to check for a fever. Symptoms can last about 7 days and the infection is contagious to others shortly before and while experiencing symptoms. It is mainly caused by four pathogens - Streptococcus pneumoniae (pneumococcal pneumonia), Haemophilus influenzae type B (Hib), influenza, and respiratory syncytial virus (RCV). Those pathogens may co-infect. Pneumococcal pneumonia is the largest cause of LRI mortality. [Wikipedia-LRI-2019], [GBD-2019-Capstone-Appendix-LRI]

The lower respiratory tract or lower airway is derived from the developing foregut and consists of the trachea, bronchi (primary, secondary and tertiary), bronchioles (including terminal and respiratory), and lungs (including alveoli). It also sometimes includes the larynx. [Wikipedia-LRI-2019]

Transmission of LRI may occur via several pathways, including direct physical contact, fomites, direct droplet spread, and suspended small particles. Intermingling of large numbers of people can facilitate transmission of respiratory pathogens. [CDC-LRI-2019]

In GBD 2016, malnutrition was identified as a leading risk factor for lower respiratory infection mortality among children younger than 5 years and, together with air pollution (both household and ambient) and increased antibiotic use, was identified as a focus for targeted intervention measures. [Lancet-LRI]

Todo

Describe more about deaths and complications due to LRI. Talk about current vaccination against influenza and pneumonia. https://apps.who.int/iris/bitstream/handle/10665/241904/WER8714_129-144.PDFp

Modeling LRI in GBD 2019

The GBD LRI model comprises a fatal and a nonfatal model. The outputs of each of these are fit to four etiologies, which are modeled separately.

LRI deaths are estimated using separate CODEm models for children under 5 and persons aged 5-95+, due to the significant difference in fatality patterns. These models run using CoD data from vital registration systems, surveillance systems, and verbal autopsy, along with a set of covariates updated slightly from those used in GBD 2017. [GBD-2019-Capstone-Appendix-LRI]

These estimates are then adjusted using CodCorrect to fit the overall mortality envelope estimated by the GBD, and mapped to etiologies by location, year, age, and sex.

The case definition used for the nonfatal LRI model is “clinician- diagnosed pneumonia or bronchiolitis”. Primary input data types include incidence and prevalence data from population surveys, scientific literature, and hospital/claims records. The modelers first adjust survey data for seasonality; then all input data with a non-reference case definition is adjusted using correction factors estimated with MR-BRT. The modelers defined time to recovery as 10 (5-15) days, which corresponds with a remission rate of 36.5 recoveries / person-year. LRI severity splits are obtained from a meta-analysis, and then the DisMod outputs are split according to severity before disablility weights for YLD calculation are applied. [GBD-2019-Capstone-Appendix-LRI] Note that as DisMod estimates an unrealistically high birth prevalence, after discussions with Theo and Nick, the modelers decided to set birth prevalence to zero.

Todo

Ask sim science, and then gbd team, what “model-MR” is. Different from dismod?

LRI viral etiologies include influenza and respiratory syncytial virus (RSV), and bacterial etiologies include Streptococcus pneumoniae and Haemophilus influenzae type B (Hib). The two types of etiologies are modeled using two different counterfactual strategies, and then for each etiology a PAF is calculated. Note that as LRI pathogens can co-infect, these PAFs can overlap. Due to a lack of data, the modelers did not map neonatal deaths to etiologies. [GBD-2019-Capstone-Appendix-LRI]

The viral etiologies were modeled using the following formula:

\[PAF = Proportion(modeled)*(1-\frac{1}{OR})\]

Here, Proportion is the proportion of LRI cases that test positive for influenza or LRI, and OR is defined to be the odds ratio of LRI given the presence of the pathogen. The odds ratios were obtained from a log-linear interpolation model, and the proportion data for each etiology was modeled using DisMod. [GBD-2019-Capstone-Appendix-LRI]

The bacterial etiologies were modeled using a vaccine probe design: the modelers first calculated the ratio of vaccine effectiveness against nonspecific pneumonia to pathogen-specific pneumonia. Estimates were adjusted by vaccine coverage and exoected vaccine performance to generate country- and year- specific PAFs. DisMod was used to model an age pattern, resulting in the final location- year- and age- specific PAF estimates. Due to a lack of vaccine efficacy data for children over two years old, the modelers did not map LRI in over-5 year olds to Hib. [GBD-2019-Capstone-Appendix-LRI]

GBD hierarchy

../../../../_images/lri_hierarchy1.svg

c_{} - cause_{gbd_id}, s_{} - sequelae_{gbd_id}

GBS stands for Guillain-Barré syndrome.

Cause Model Diagram

../../../../_images/lri_disease_model1.svg

Model Assumptions and Limitations

Because DisMod estimated an unrealistically high birth prevalence, the modelers set birth prevalence to zero. Consequently, the birth prevalence, incidence, and prevalence available from get_outputs are incongruous with one another.

This model is designed to be used for estimating DALYs due to LRI that are averted from a country-level intervention(e.g. food fortification or supplementation given to a percentage of the population) that can reduce LRI incidence as a downstream effect.

There is substantial additional effort in GBD to divide LRI burden into the aetiologies of LRI, but we do not include aetiologies in this simple model.

There are three sequelae associated with LRI, including moderate LRI, severe LRI, and Guillain-Barré syndrome due to LRI. We are not tracking the long-term effects of Guillain-Barré syndrome (which can include paralysis, for example). However, since the prevalence of GBS is so low, there would likely not be great benefit in capturing its long-term YLDs in addition to its short-term YLDs.

Todo

Describe more assumptions and limitations of the model.

Data Description

Definition

State

State Name

Definition

S

Susceptible

Susceptible but does not currently have LRI

I

Infected

Currently infected and having the condition
States Data

State

Measure

Value

Notes

S

birth prevalence

0

S

prevalence

1-prevalence_calculated

S

excess mortality rate

0

S

disability weights

0

I

birth prevalence

0

I

prevalence_calculated

For early neonatal age group: (birth_prevalence_I + (incidence_rate_c322 * duration_c322))/2. For all other age groups: incidence_rate_c322 * duration_c322

Justification included below. Early neonatal age group exception due to non-steady state dynamics in this age group given birth prevalence of zero causes increasing prevalence within age group and short duration of age group. Citation on these dynamics and approximations here for reference.

I

excess mortality rate

\(\frac{\text{deaths\_c322}}{\text{population} \,\times\,\text{prevalence\_calculated}}\)

I

disability weights

disability_weight_s670 \(\times\) prevalence_s670+ disability_weight_s669 \(\times\) prevalence_s669 + disability_weight_s671 \(\times\) prevalence_s671

ALL

cause specific mortality rate

\(\frac{\text{deaths\_c322}}{\text{population}}\)

We calculate prevalence using the equation prevalence = incidence * duration. (See assumptions and limitations for the need to replace GBD’s prevalence). This is appropriate because LRI has a short and relatively uniform duration of 7.79 days (95% CI 6.2–9.64) days [GBD-2019-Capstone-Appendix-LRI]. This assumption is valid under steady state conditions. However, the prevalence of LRI is not in steady state for the early neonatal age group given a birth prevalence of 0 and a short duration of the age group (prevalence will increase as the population ages within the age group). Therefore, we calculate the prevalence in the early neonatal age group as an average of the birth prevalence and the approximated prevalence under a steady state transition (incidence * duration). This is approach is discussed in this citation for reference.

Transition Data

Transition

Source

Sink

Value

Notes

i

S

I

\(\frac{\text{incidence\_rate\_c322}}{(1-\text{prevalence\_calculated})}\)

Incidence in GBD are estimated for the total population. Here we transform incidence to be a rate within the susceptible population.

r: USED IN CIFF AND IV IRON SIMULATIONS AS WELL AS MODELS 1-11 OF NUTRITION OPTIMIZATION SIMULATION

I

S

(-1/time_step)*log(1-time_step/duration_c322)

Where time_step is the duration of the simulation time_step in years. Use the np.log() function. See discussion of this equation on the diarrheal diseases cause model document and Choosing an Appropriate Time Step page page for more information.

r: FOR USE IN NUTRITION OPTIMIZATION SIMULATION AFTER IMPLEMENTATION OF VARIABLE TIMESTEPS

I

S

1/duration_c322
Data Sources

Measure

Sources

Description

Notes

birth_prevalence_c322

como

0

No birth prevalence

prevalence_calculated

Calculated from incidence (como) and duration (literature/gbd)

Duration-based calculation of LRI Prevalence

deaths_c322

codcorrect

Deaths from LRI

population

demography

Mid-year population for given age/sex/year/location

incidence_rate_c322

como

Incidence rate of LRI within the entire population

remission_rate_m1258

dismod-mr

Remission rate of LRI within the infected population

disability_weight_s{sid}

YLD Appendix

Disability weights associated with each sequela

Note Guillain-Barre due to LRI is included in sequelae.

prevalence_s{sid}

como

Prevalence of each sequela with id ‘sid’

duration_c322

(7.79 days; 95% CI 6.2–9.64; normal distribution of uncertainty)/365

Mean duration of an LRI case (in years). From the YLD appendix

This value should not vary by age group despite the fact that the duration is longer than the length of the early neonatal age group.
Restrictions

Restriction type

Value

Notes

Male only

False

Female only

False

YLL only

False

YLD only

False

YLL age group start

Post neonatal (age group ID 4, 1 month to 1 year)

GBD age group start is early neonatal (age group ID 2, 0-6 days)

YLL age group end

Age 95+

GBD age group id is 235

YLD age group start

Post neonatal (age group ID 4, 1 month to 1 year)

GBD age group start is early neonatal (age group ID 2, 0-6 days)

YLD age group end

Age 95+

GBD age group id is 235

Note

A note on the LRI age start parameter:

This Vivarium modeling strategy sets the LRI cause model age start to the post neonatal age group (1 month to 1 year) despite the GBD age start parameter being the early neonatal age group (0 to 6 days). We exclude the early and late neonatal age groups from the cause model as a strategy that allows us to increase the timestep of our cause models.

The rationale behind this modeling decision is related to the Relationship between timesteps and modeled rates in Vivarium as described on the Choosing an Appropriate Time Step page that is exacerbated by the inclusion of the low birth weight and short gestation risk factor in the model. Essentially, because the LBWSG risk factor affects LRI excess mortality rates in our models during the neonatal age groups and the LBWSG relative risk values for the highest risk categories are quite large (up to 700!), the inclusion of the LBWSG risk effects on LRI causes individual-level LRI excess mortality rates to be too large to accurately approximate in our models without a very small timestep, which leads to underestimation of neonatal LRI mortality rates with a timestep on the order of 0.5 days.

Therefore, we employ the following strategy:

  • Model the LRI SI cause model as described in this document for ages older than late neonatal only, and

  • Include LRI as an unmodeled cause that is affected by the LBWSG risk factor (see the LBSWG risk effects page for details). This will allow us to model LRI CSMR rather than EMR among the neonatal age groups, which is lower in magnitude and therefore less easier to approximate at larger simulation timesteps. Notably, this strategy does not allow us to model years lived with disability due to LRI among the neonatal age groups.

This strategy allowed us to increase the simulation timestep to 4 days and still meet verification criteria.

Validation Criteria

Baseline vivarium model results should compare to GBD artifact data with respect to age-, sex-, location-, and year-specific LRI:

  • Prevalence

  • Incidence rate

  • Remission rate

  • Cause-specifc mortality rate

  • Excess mortality rate

  • YLDs due to LRI

  • YLLs due to LRI

Note

The prior bound for the LRI remission rate is 7.3 days, which is longer than the duration of the early neonatal age group (6 days), so theoretically there should be few or no remitted cases of LRI in the early neonatal age group. However, LRI birth prevalence is expected to be greater than LRI prevalence in the early neonatal age group due to LRI’s excess mortality rate.

References

[Wikipedia-LRI-2019] (1,2)

Lower respiratory tact infection. From Wikipedia, the Free Encyclopedia. Retrieved 22 Nov 2019. https://en.wikipedia.org/wiki/Lower_respiratory_tract_infection

[CDC-LRI-2019]

Respiratory Infections (The Yellow Book). Centers for Disease Control and Prevention, 2019. Retrieved 20 Dec 2019. https://wwwnc.cdc.gov/travel/yellowbook/2020/posttravel-evaluation/respiratory-infections

[Lancet-LRI]

The Global Burden of Lower Respiratory Infections: Making Progress, but We Need to Do Better (Volume 18). The Lancet Infectious Diseases, 2018. Retrieved 20 Dec 2019. https://www.sciencedirect.com/science/article/pii/S1473309918304079?via%3Dihub

[GBD-2019-Capstone-Appendix-LRI] (1,2,3,4,5,6,7)

Appendix to: GBD 2019 Diseases and Injuries Collaborators. Global burden of 369 diseases and injuries in 204 countries and territories, 1990–2019: a systematic analysis for the Global Burden of Disease Study 2019. The Lancet. 17 Oct 2020;396:1204-1222