Quantitative Comparison of Human Health Risks and Benefits Caused by Animal Antibiotic Use Tony Cox

1
Quantitative Comparison of Human Health Risks and
Benefits Caused by Animal Antibiotic Use Tony
Cox Cox AssociatesSRA SymposiumSeptember 29,
2004
2
Goals

• Predict changes in frequency and severity (
  risk) of human illnesses that will be caused by
  proposed changes in animal antibiotic uses
  (AAUs).
• Desired outputs Predicted changes in illnesses,
  illness-days, QALYs lost, etc. per year (for
  population risks) and per capita-year (for
  individual risks)
• Changes in frequencies and severities of
  illnesses relevant human health consequences
  of interventions
• Identify interventions that most reduce human
  illnesses and treatment failures
• Quantify uncertainties about true human health
  consequences of changing AAUs

3
Q How to do it?

• A Rapid Risk Rating Technique (RRRT) approach
• Change in risk (change in exposure) x (average
  human health risk per unit of exposure), summed
  over paths
• Change in exposure reflects how changing AAU
  affects microbial loads in ingested servings
• Average risk per unit exposure illnesses (or
  illness-days, deaths, QALYs lost, etc.) per unit
  of exposure
• E.g., ? Risk (servings ingested per year) x (?
  illnesses per serving) x (QALYs lost per illness)
  
• So, ?Risk ?(Exposure x Dose-Response factor x
  Consequence factor)
• Sum over paths (subpopulations, bacteria,
  strains, commodities) with different values of
  these factors

4
Rapid Risk Rating Basic Logic

• ? animal antibiotic use (? AAU)
• ?
• (? animal illness rates) (? animal microbial
  loads)
• (consider both resistant and susceptible
  bacteria)
• ?
• human exposure (e.g., CFUs per serving)
• ?
• covariates ? ? human illness rates ? ? QALYs/yr.

5
Rapid Risk Rating Resistance

• ? animal antibiotic use (? AAU)
• ?
• (? animal illness rates) (? animal microbial
  loads)
• (both resistant and susceptible microbes)
• ?
• human exposure (e.g., CFUs per serving)
• ?
• covariates ? ? human illness rates ? ? QALYs/yr.
• ? ?
• prescriptions ? ? human resistance rates

6
Rapid Risk Rating Sub-models

• ? animal antibiotic use (? AAU)
• ? animal health model release
• (? animal illness rates) (? animal microbial
  loads)
• (both resistant and susceptible microbes)
• ? exposure model
• human exposure (e.g., CFUs per serving)
• ? dose-response model
• covariates ? ? human illness rates ? ? QALYs/yr.
• ? consequence model ?
• prescriptions ? ? human resistance rates

7
Rapid Risk Rating Quantitative Factors

• ? animal antibiotic use (?AAU)
• ? ?(AS)/(?AAU) exposure, part 1
• (? animal illness rates) (? animal microbial
  loads)
• (both resistant and susceptible microbes)
• ? (?CFUs per serving)/(?AS) exposure 2
• human exposure (e.g., CFUs per serving)
• ? (?illnesses)/(?CFU) dose-response
• ?covariates ? ? human illness rates
• ? (?QALYs)/(? illness) consequence
• ? prescriptions ? ? QALYs lost/yr. ? ? resistance

8
Example Exposure factor calculation

• Predict change in CFUs per serving as follows
• (?AS) x (CFUs/AS serving) (CFUs/AS
  serving).
• ?AS change in fraction of servings that come
  from infected animals (or AS flocks)
• AS serving serving from infected animal
• Predict changes in animal illness rates (?AS)
  and resistance rates (? fraction resistant) in
  short and long runs (e.g., from 1 to 1.2 from
  0 to 20 to 0, etc.) caused by ?AAU.
• Use European data, statistics on usage rates and
  prevention efficacies, and/or infectious disease
  predictive models

9
Example of Exposure Calculation

• If
• Proposed change causes ?AS to increase 1, and
• CFUs per AS serving 10 x (CFUs per AS
  serving)
• Then
• Predicted increase in average CFUs/serving is
  (0.01) x (10 1) x (CFUs per AS serving)
• If baseline AS 0 (so AS 100), then
  Predicted New CFUs per serving (1.09) x
  (Current CFUs per serving)
• Sensitivity If microbial load ratio is 2
  instead of 10, then New (1.01) x (Current)
  instead of 1.09x.

10
Quantifying Dose-Response and Consequence Factors

• Dose-response factor (illnesses per serving) or
  (illnesses per CFU) at relevant exposure level
• Linear no-threshold model provides simplest base
  case
• Nonlinear models provide sensitivity analyses
• Log-exponential model
• Mixture-of-Beta-Poisson models
• Heterogeneous subpopulations (finite mixure)
  models
• Consequence factor QALYs lost per illness
  (fraction resistant)(QALYs per resistant case)
  (1 fraction resistant)(QALYs/susceptible case)
• Bounded uncertainty 0 ? fraction resistant ? 1

11
Dealing with Uncertainties

• Some uncertainty modeling approaches
• Use conservative bounds for unknown quantities
• Refine with probabilities
• Multiplicative uncertainty factors (e.g.,
  log-normal uncertainty factors for products)
• Distributions (if known) and Monte Carlo.
• Perform sensitivity analyses
• Conservative bias Choose bounds and estimate
  each uncertain input quantity so that conclusions
  are likely to be strengthened by further data.

12
RRRT Framework Main Ideas

• ? Population Risk (e.g., in illness-days/year)
• (? use)(? exposure /? use)(?illnesses/?exposure
  ) (? health consequences/? illness)
• ?Exposure x dose-response x health consequence
• Sum over groups, strains, outcomes, uncertainties
• All ?s are causal, not association-based/attribut
  able
• (Can use conditional distributions and Bayesian
  Network or causal graph algorithms instead of
  multiplying ratios)
• Designed to be simple, correct,
  flexible/realistic
• Work needed estimating and documenting factors
  (and bounds, distributions, or uncertainty
  factors)

13
Examples of RRRT Calculations

• Calculate plausible upper bounds on current risks
  that might be prevented by stopping current
  animal antibiotic uses
• Virginiamycin/QD and VREF example
• Preventable fraction ? attributable fraction
• Calculate plausible lower bounds on avoidable
  increase in current risks that might be caused by
  stopping current animal antibiotic uses
• Macrolide and campylobacter example
• Compare upper-bound estimates of risk prevented
  to lower-bound estimates of risk caused.
• Rational (consequence-driven) decision-making
  requires considering both.

14
Example RRRT Risk Calculation for QD
15
Example of RRRT Logic for QD

16
Comments on RRRT approach

• Transparent logic (multiplication)
• Technical and policy assumptions can be
  explicitly identified/debated
• Upper bounds avoid non-crucial debate /
  uncertainty
• Remaining debate/deliberation is apt to be
  informative and useful clear roles for data
• Factors based on published data and explicit,
  verifiable calculations
• Allows short-term and long-term impact estimates
• Need population dynamics sub-model for long term
• Can stop calculation at any point with an
  estimated upper bound on final answer
• (Monte Carlo and log-normal uncertainty analyses)

17
Macrolide-Campylobacter RRRT Example

• Health RISK from current use Adverse effects
  (e.g., QALYs lost) per year caused by resistance
  from macrolide use in animals (mainly poultry)
• Health BENEFIT from current use Adverse effects
  prevented by reduced illness rates in people from
  reduced animal variability/illnesses
• Which is larger, RISK or BENEFIT?
• Are RISK and BENEFIT large enough to be worth
  worrying about?
• RRRT provides quick approximate answers.

18
Example RRRT Calculation Macrolides and
Campylobacter
19
RRRT Example Calculations (Cont.)
20
RRRT Example Calculations (Cont.)
21
Summary of RRRT Risk Estimate

• RISK The expected human health harm that could
  be prevented by eliminating macrolide-resistance
  in campylobacter is lt 1 treatment failure
  (causing perhaps 0-2 excess days of severe
  campylobacteriosis duration) per year in the US.
  
• This estimate may be too high or too low by a
  factor of about 18 based on estimated uncertainty
  factors.
• Key uncertainty not quantified What are the
  clinical benefits of macrolide therapy for campy?
• How important are indirect paths?

22
RRRT Health Benefits Calculation
23
RRRT Benefits Calculation (Cont.)
24
Main Results

• Current use of macrolides in chickens is expected
  to have human health BENEFITS gt 1,000 times human
  health RISK (40,000 vs. 2 6658 vs. 1)
• Withdrawing AAU may do much more harm than good
  to human health ? High VOI from finding out!
• Conclusion is robust to many model uncertainties
• Key uncertainties Microbial load ratios,
  dose-response relation, animal illness increase
  if macrolide use withdrawn (possibly gtgt 0.5)
• Using FDA-CVM dose-response relation (nonlinear)
  increase estimated benefit of macrolide use
  14-fold.

25
Other Considerations

• Some omitted factors may increase estimated
  benefits further
• Behaviors Antibiotic substitution (Casewell et
  al., 2003)
• AAU synergies in suppressing animal illness rates
  (e.g., combined prevention control)
• Other animal illnesses (e.g., necrotic enteritis)
• Other food animal-borne pathogens (e.g.,
  Salmonella)
• Future dynamics Reduced need to treat human
  patients may reduce human-use selection pressures
  and resistance
• Opportunistic infections could increase
  resistance harm
• No empirical evidence Effect too small to
  detect easily
• Uncertain co-selection/cross-resistance impacts
  bounded
• lt 0.1 case/year, bound from previous VM/QD RRRT
  assessment
• Other animal food pathways are relatively minor

26
Conclusions on RRRT

• RRRT quantifies likely human health impacts (good
  and bad) of proposed changes in AAUs.
• Uses available data to bound risk estimates
• Puts bounds on major uncertainties (set fractions
  to 1)
• Can extend to include time, Monte Carlo
  simulation
• RRRT model fragments can be re-used for other
  antibiotics/bacteria/situations.
• Allows remaining disagreements/uncertainties to
  be clarified and addressed with data.
• Conforms to Guidance 152 general approach, but
  exceeds it by estimating actual health impacts

27
Possible futures for risk assessment

• Must become more realistic/accurate to survive!
• Objective/data-driven vs. subjective/judgment-driv
  en
• Participatory vs. command-control risk management
• Prediction and prevention vs. attribution and
  blame
• Decision/change/consequence oriented vs.
  reaction/situation/attribution oriented
• Consider full health consequences of actions
  (e.g., on susceptible and resistant bacteria in
  patients) vs. consider only selected consequences
  (e.g., resistant bacteria in animals and healthy
  people).

28
Using Quantitative Risk Assessment (RA) to
Improve RM Decisions

• Promise of quantitative risk assessment (RA)
• Better human health outcomes better bets
• Science-based decisions (more consequence- and
  fact-driven less expert judgment/situation-driven
  )
• More data-driven, participatory, open, and
  interpretable than qualitative RA
• Tiered approach (qualitative/quantitative) may be
  effective if qualitative estimates are not too
  misleading
• Quicker, easier, cheaper, more useful, more
  valid/accurate than qualitative assessment.

29
Comments on RRRT Method

• Same approach for estimating health benefits and
  health risks of AAU.
• Rigorous mathematical foundations
• Sum of random number of random variables
• Compound Poisson approximation for sporadic
  events
• Conditional probabilities justify product form
• Bayesian belief network
• Zero exposure implies zero risk. Zero
  consequence implies zero risk. (Unlike Guidance
  152 labels)
• Can distinguish significantly different risks

30
Russell 2003 Data