Quantitative Risk Analysis for Animal Antibiotics Franz Edelman Award Competition

Quantitative Risk Analysis for Animal Antibiotics Franz Edelman Award Competition May 1, 2006 The team • Problem description: Rich Carnevale, Vice President of Regulatory, Scientific and International Affairs, AHI • OR solution: Doug Popken, Director, Simulation Modeling, Cox Associates • Explanation and impact: Tony Cox, President , Cox Associates $100 million question How to quantify the true human health effects of using animal antibiotics? – No hype – No panic – Credible facts Which intuition is correct? • Optimistic: Prudent use causes… – Better animal health – Less bacteria in food – Better human health • Pessimistic: Prudent use causes… – Resistant bacteria in animals – Treatment failures in humans – Harm to human health Fear: Animal antibiotics are killing patients! Fact: Documented cases = 0 Is zero too high? • Zero cases found ≠ zero risk • Zero risk ≠ zero perceived risk • Groups prey on fear: ???? “ During 2005, Keep Antibiotics Working helped generate more than 1,500 print, radio and TV stories, editorials, letters to the editor and op-eds, reaching a combined audience of more than 49 million people.” AHI Role Politicians Farmers Physicians TRUTH? Media Industry Veterinarians Activists Regulators Manufacturers Facts/data/science/OR? What difference did OR make in just four years (2000-2004)? Fear builds 1999: 1999: 2000: 2001: 2002: 2003: 2003: National Research Council: "The risk is greater than zero, but basically incalculable“ “ Livestock drug could ruin human antibiotics” “ Worries rise over effect of antibiotics in animal feed ” FDA considers banning virginiamycin "[H]ow long would it be before [Synercid] was dead?“(from The Killers Within) Editorials: “ antibiotic use in food Cut animals” Congress: "Preservation of Antibiotics for Medical Treatment Act" proposes bans “ Germs in meat stir debate on drug use in livestock” “ BOSTON (AP) October 17, 2001 – [CDC researchers] found that more than half of 407 supermarket chickens… carried the sometimes fatal germ Enterococcus faecium in a form resistant to Synercid, one of the few drugs of last resort against the infection.” “ the study, 1% of human stool samples also In carried the resistant germ. But the researchers warned of future increases and suggested that use of the livestock drug, virginiamycin, may need to be limited.” AHI’ OR focus s • How would changing virginiamycin use change human health risks? – Shift focus from blame to change – Shift focus to likely results, not intentions • Veterinary intuition: Virginiamycin protects animals… and therefore people • Animal use does not harm people – Different strains of bacteria – Not easily transferred European ban -- oops! Illness and resistance rates in patients increased significantly after precautionary bans. Hayes and Jensen, 2003 AHI’ goals for OR modeling s • How big a risk from continued use? – What is likely to be true? – What is definitely true? – Simulation models + sensitivity analyses – Systems dynamics models – Use conservative bounds to close uncertainty gaps • Explain/communicate it… simply and clearly! Main Concern Could VM in the food supply… … create antibiotic-resistant strains of generally harmless bacteria… … which could threaten the immune-compromised? OR Modeling • Farm-to-Fork model surprise, 2001: – Discrete-event simulation predicted that reducing animal antibiotics increased harm to human health! – But are the results credible? • New “ Clinic-to-Farm”approach, 2002 – Upper-bounding approach using biomarker data let us avoid many scientific uncertainties: doseresponse function, subpopulations, etc. • Bayesian Monte Carlo analysis, 2003 – Future “ resistance epidemic”has zero probability • Rapid Risk Rating Technique, 2004 – Simplified explanations; transparent analysis Simulation models Informative but not necessarily credible… Simulation models of microbial loads and health effects give insight… but are too detailed for ready acceptance Farm-to-Fork Simulation: Summary • Generated insights and testable hypotheses: • Halting antibiotic use increases variability in animal weights and health • Increased variability increases right tail of microbial load distribution in meat servings → Increases human health risk – Especially for strictly convex dose-response with results robust to many uncertainties – Using non-linear dose-response increases predicted human health harm from ban 14-fold • Interesting insights, but nothing proved. • Value-of-information perspective: Worth finding out! Revelation from simulation: Unhealthy animals are the main threat to human health Withdrawing animal antibiotics shifts microbial load distribution rightward. Higher loads → more sick people. “ Black box”challenge Clear logic does not prove numbers in simulation are correct! What about failures of imagination? Can’ simulate what’ not in t s the model. So, how can we ever trust that the results have not overlooked something crucial? Living bacteria are notoriously adaptive and tricky! Answer 1: Clinic-to-Farm Modeling • Start with all cases of illnesses per year – Normal distribution fit to US data (thousands of cases/yr) – Clusters of Markov chains (“ mixture distribution” ) fit to Australian data (a dozen or fewer cases/yr) • Estimate maximum fraction of illnesses leading to treatment failures using: – Historical resistance – Prescription – Clinical success rates Clinic-to-Farm (cont’ d.) • Determine maximum fraction that could have been caused by virginiamycin used in animals – Estimated from molecular biomarker data (esp gene) – Assume all are caused by animal use and would be prevented by a ban (upper bounding) • Chain of conservative calculations for each component: – Approximate unknown relations with bounds – Choose bounds biased against main conclusions Answer 2: Bayesian Monte Carlo Uncertainty Analysis Challenge: Quantify human health risk of future resistant illnesses, given decades of exposure but < 1% resistance in US. • Bayesian Monte Carlo (cont’ d.) • Convert widely accepted system of ordinary differential equations to stochastic transition model • Refine estimated transition parameters – Condition on historical data using published values as means of log-normal priors – Assume resistance levels above detection threshold are detected – Simulate system over time – Reject input combinations producing resistance levels above historical levels – Distribution of accepted combinations → posterior distribution of transition parameters → risk prediction Bayesian Monte Carlo results: There’ nothing to fear s • Quantitative risks < 1 × 10-6 even for most-threatened (ICU patient) population. • Societal risks < 1 excess death/year for whole U.S. OR Modeling • Farm-to-Fork model surprise, 2001: – Discrete-event simulation predicted that reducing animal antibiotics increased harm to human health! – But are the results credible? • New “ Clinic-to-Farm”approach, 2002 – Upper-bounding approach using biomarker data let us avoid many scientific uncertainties: doseresponse function, subpopulations, etc. • Bayesian Monte Carlo analysis, 2003 – Future “ resistance epidemic”has zero probability • Rapid Risk Rating Technique, 2004 – Simplified explanations; transparent analysis Assessing current risks from continued use… credibly! Attribution approach to quantifying risk (Rapid Risk Rating Technique, RRRT) Multiply current total VRE cases per year by: 1. 2. 3. 4. 5. 6. 7. 8. Potentially treatable fraction (VREFA fraction) Fraction that are not nosocomial Fraction that could come from chicken Fraction prescribed human antibiotic (Synercid) Fraction of patients that are Synercid-tolerant Fraction of infections that are Synercid-resistant Fraction of resistance that would be prevented by discontinuing animal antibiotic (virginiamycin) use Fraction that result in harm (× QALYs lost/case) Example RRRT Risk Calculation Factor VRE × % VanA VREF fraction not nosocomial fraction from chicken fraction given Synercid fraction Synercidresistant effective treatment fraction if no resistance fraction prevented if virginiamycin withdrawn Excess mortalities/case RISK = Product of above Mean 40,000 × 0.61 0.17 0 to 0.124 < 0.07 0 to 0.01 0.72 <1 0-0.15? < 0.04/year Data Eliopoulos ‘ ; Jones ‘ 98 95 Bischoff ‘ Austen ‘ 99; 99 Willems, 2000, 2001 (all paths) linezolid failure rate Eliopoulos 1998 Jones et al, 1999 Moellering et al., 1999; Linden, 2002 Upper bound. (Value may be zero, based om Europe) Linden 2002, may be zero 40000*.61*.17*.12* .07*.01*.72*.15 = .04 Comments on RRRT approach • Transparent logic (multiplication, conditioning) – Technical, policy assumptions explicitly identified – Upper bounds avoid non-crucial debates • Factors based on published data and explicit, verifiable calculations • Allows short- and long-term impact estimates – Population dynamics sub-model • Can stop calculation at any point with an estimated upper bound on final answer • Simple sensitivity analyses RRRT unaffected by unknowns • Beginning with total cases takes into account many hard-to-model unknowns – Cross-contamination – Secondary transmission and amplification – Unknown environmental paths, etc. • Imagination not necessary How harmful is a ban? • A similar approach can estimate the human health harm from a ban • Harm from ban = health benefit lost if use ceases RRRT benefits calculation US population Average fresh chicken servings per capita-year Current average risk of campylobacter per serving 292E6 (US Census) 38 (FDA-CVM) 1E-5 = (total cases per year) × (10% fraction from chicken)/(total servings) Increased illness fraction > 0.5% ? (~17% in Norway after in chickens if no ban for NE) virginiamycin Excess risk per serving from ill flock (e.g., necrotic enteritis positive, NE+) Campylobacteriosis cases prevented per year: > 6658 ≥ 1.2E-4 [bound from linear doseresponse model, microbial load ratio = 10 (Russell, 2003)] 292E6*38*0.005*1.2E-4 = 6658 (For comparison, RISK < 0.04/yr.) Benefits may dwarf risks • Potential human health benefits from reduced illness risk is more than 10,000 times the potential risk from increased resistance and treatment failures • 40,000 illness-days (≈ 6648 × 6 days/illness) and 40 serious campylobacteriosis cases prevented per year vs. < 1 potential excess treatment failure per year • Withdrawing VM may do more harm than good to human health → High VOI from finding out! Results stand unchallenged The RRRT risk numbers and risk conclusions have not been challenged! – FDA, CDC, national and international scientific audiences usually commented favorably – Multiple researchers and WHO have referenced and adopted the ideas and approach: Snary et al., 2004; Bywater, 2005; McDermott et al., 2005 (FDA); Singer et al., 2005; Hurd et al., 2006, Wikipedia, etc. – The benefits/value-of-information perspective is not so widely accepted (yet) OR impact • Risk quantified: 0 to < 1 statistical life saved over 5 years in Australia and US combined by a virginiamycin ban… and decreasing rapidly. – Cox & Popken, Quantifying human health risks from virginiamycin used in chickens. Risk Analysis 2004 Feb;24(1). • Australia and US allow continued use of virginiamycin. – November 2004: FDA agrees quantitative risk from use is at most very small (if not zero). Phibro spreads the news (November 30, 2004) “ Phibro Animal Health Corporation announced today that the Center for Veterinary Medicine (CVM) of the Food and Drug Administration (FDA) has released the findings of its draft risk assessment for the animal drug virginiamycin. In the FDA's draft report, the authors demonstrate that the continued use of virginiamycin –a medicated feed additive manufactured and marketed by Phibro Animal Health – in livestock and poultry feed poses no significant risk to human health.” Philbro: Consistency, health ” The findings are consistent with our own … conclusions about the historical safety of virginiamycin and its role in protecting both animal and human health." “ The FDA findings are consistent with extensive analysis conducted… by Dr. Tony Cox, one of the world's leading authorities on risk assessment.” “ Dr. Cox not only determined that the continued … use of virginiamycin poses minimal risk to human health, but that a ban could lead to a significant increase in food-borne illness. By improving the health of livestock and poultry, virginiamycin minimized the likelihood of human exposure to harmful food-borne bacteria.” FDA Statement, January 6, 2006 Real-world effect of no bans? USDA Release No. 0128.05, April 14, 2005: • “ report [by CDC, USDA, and FDA] showed A important declines in foodborne infections due to common bacterial pathogens in 2004.” – “ From 1996-2004, the incidence of E. coli O157 infections decreased 42 percent.” – “ Campylobacter infections decreased 31 percent, – Cryptosporidium dropped 40 percent – Yersinia decreased 45 percent – Overall, Salmonella infections dropped 8 percent… “ • Most resistance rates for domestically acquired infections remained stable or decreased $100 million answer • Got science back on the table • Animal and human health, not or human health • Member companies applied new methods to – Virginiamycin (Phibro) – Macrolides (Elanco) – Baytril for cattle (Bayer); policy (Alpharma); etc.. • Cited and used by companies and health risk researchers worldwide. • No animal antibiotic growth promoters have been banned in the US! • > Hundred million dollar/yr. market saved • Non-essential usage is down due to market forces Lessons learned • OR helped everyone focus on probable consequences rather than on fears – Focusing on consequences helps avoid unwanted results from precautionary approach • To bridge gaps in knowledge and data credibly: – Use data-driven totals and conservative bounds to estimate maximum preventable fractions. – Simple, robust, data-driven calculations are credible – Sensitivity analyses, systems dynamics models, simulation models, Monte Carlo uncertainty analyses, etc., add comfort Portability • RRRT “ template”already successfully applied to macrolides (Elanco), enrofloxacin for cattle (Bayer), as well as virginiamycin (Phibro) • Basic approach has recently been applied successfully to bound preventable fractions for – Exposure-related lung cancer (Cox & Sanders ‘ 06) – Gene flows from herbicide-resistant grasses – Both use biomarker data among total cases at risk to obtain useful bounds on size of risk For everyone… Safer, cheaper food Better health Confidence, not panic Thank you! Quantitative Risk Analysis for Animal Antibiotics Questions and answers