the impact of quadrivalent influenza vaccine (qiv) in canada: some insights from a dynamic model ed...
TRANSCRIPT
The impact of quadrivalent influenza vaccine (QIV) in Canada: Some insights from a dynamic modelEd Thommes, PhDHealth Outcomes ManagerGlaxoSmithKline Canada &Department of Mathematics & Statistics, University of Guelph
4Strain dynamic influenza model team:
Chris Bauch
Professor, Dept. of Applied Mathematics
University of Waterloo, ON
Geneviève Meier
Director, Health Economics, Vaccines
GlaxoSmithKline
Wavre, Belgium
Ayman Chit
Director, Health Outcomes and Economics North America
Sanofi Pasteur
Toronto, ON
Afisi Ismaila
Director Therapy Area
GlaxoSmithKline
Research Triangle Park, NC, USA
2
Outline
• Background: What is QIV?
• Overview of the 4Strain dynamic transmission model
• Calibrating the influenza “natural history” input parameters
• Test case: Ontario’s adoption of universal influenza immunization
• TIVQIV switch results: outcomes prevented and cost-effectiveness
• Summary
3
Background: TIV
Current trivalent influenza vaccines (TIV) contain 2 influenza A virus types: H3N2, H1N1 and one influenza B lineage
Annual strain recommendation is based on surveillance Recommended strains may not reflect current
circulating strains
Co-circulation of B/Victoria and B/Yamagata
Background: Influenza B
Two main genetic lineages in circulation:
1. Victoria (1987)
2. Yamagata (1988)
B Victoria and B Yamagata have co-circulated in recent years
Mutation rate is slower compared to influenza A viruses
Vaccine mismatch for influenza B: Canada
Adapted from Fluwatch http://www.phac-aspc.gc.ca/fluwatch/ and NACI http://www.phac-aspc.gc.ca/naci-ccni/
20
00
-01
20
01
-02
20
02
-03
20
03
-04
20
04
-05
20
05
-06
20
06
-07
20
07
-08
20
08
-09
20
09
-10
20
10
-11
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Mismatch Match
Season
Infl
ue
nza
B:
% t
ota
l ch
ara
cte
ris
ed
infl
ue
nza
is
ola
tes
GSK’s QIV: FluLaval® Tetra
Quadrivalent split-virion, inactivated influenza vaccine
Authorized for use in Canada Feb 6, 2014
Manufactured in Sainte-Foy, Quebec
A menagerie of modeling approaches…
flu model
static
tree Markov
dynamic
compartmental
individual or “agent”-based
(ABM)
8
Model structure:i) Simple S(usceptible)I(nfected)R(ecovered) model
infection
natural immunity
natural immunity waning
� ܫ �
Model structure:ii) Adding a second strain
infection
natural immunitynatural cross-protection
natural immunity waning
�ଵ�ଶ ଵ�ଶܫ �ଵ�ଶ
�ଵ�ଶ
�ଵܫଶ
�ଵ�ଶ ଵ�ଶܫ
�ଵܫଶ
•Approach of Castillo-Chavez et al. (1989)•Introduces cross-protection into model dynamics•Immunity waning: each strain sequentially, i.e.. R1R2→S1R2→S1S2 orR1R2→R1S2→S1S2
infectionvaccinationnatural immunitynatural cross-protectionvaccinated immunity waningnatural immunity waning
�ଵ �ଶ
�ଵ �ଶ�ଵ �ଶ
�ଵ �ଶ ଵܫ �ଶ �ଵ �ଶ
ଵܫ �ଶ �ଵ �ଶ
�ଵ �ଶ
�ଵܫଶ
�ଵ �ଶ ଵܫ �ଶ
�ଵܫଶ�ଵܫଶ
�ଵ �ଶ
•Success/failure determined at time of vaccination: Let ε1, ε2 be the efficacies. Then, e.g. for a person in S1S2, possible outcomes of vaccinating, are, with probability P:
• P=ε1 ε2: go to V1V2
• P= ε1(1- ε2): go to V1S2
• P= ε2(1- ε1): go to S1V2
• P=(1- ε1)(1- ε2): stay in S1S2
•Waning of vaccinated immunity occurs analogously to waning of natural immunity
NOTE: We assume that the natural immunity always lasts at LEAST as long as vaccine-conferred immunity. Thus, e.g., successfully vaccinating someone in compartment R1S2 against strain 1 has no effect
Model structure:iii) Adding vaccination
Calibrating the model to real-world data
– Ideally, we’d like to use a given region’s influenza surveillance to calibrate model parameters
– Problem: influenza surveillance very incomplete instead, used Turner et al. (2003) HTA: Calculates unvaccinated (“natural”) attack rate of influenza from placebo arms of vaccine & antiviral RCTs
– advantage of natural atk rate: Only indirectly (through herd immunity) depends on vaccination state of population
(or: avoiding “Garbage In – Garbage Out”)
12
Our calibration approach: Approximate Bayesian computation (ABC)
13
Fitting simulations: Influenza in the US, 1998-2008
influenza Ainfluenza B
Thommes et al., Vaccine, submitted
Testing the model: Ontario’s adoption of a universal influenza immunization program (UIIP)
– Implemented in 2000; world’s first large-scale universal influenza immunization program
– Resulting changes in both vaccine uptake and influenza-associated events have been studied in detail (Kwong et al., PLoS Medicine 2008). Events considered:– doctor’s office (GP) visits – emergency room (ER) visits– hospitalizations– deaths
– Objective: Assess how well our model agrees with Kwong et al.’s results
Testing the 4Strain model on Ontario’s UIIP:
Kwong et al. (2008)4Strain dynamic model
Thommes et al., Vaccine, submitted
Result: Model is overall conservative relative to Kwong et al. (2008) in predicting outcomes averted by UIIP
Relative rate ratio: ReductionOntario
=-------------------------- ReductionCanada
Result: Canada-wide TIVQIV switch brings about clear reduction in outcomes
# simulations
outcomes per season, TIV and QIV
outcomes prevented per season by QIV
influenza cases(50k-300k prevented)
GP visits(20k-120k prevented)
ER visits(1000-8000 prevented)
hospitalizations(500-4000 prevented)
deaths(50-800 prevented)
Sensitivity analysis: QIV highly cost-effective across all plausible inputs
Limitations:
– Vaccine uptake extrapolated below 12 yrs in most provinces
– Using mostly US attack rates in model calibration– Very little information about duration of vaccine-conferred
immunity to influenza (we assume 1 yr on average)– No healthy vs. at-risk stratification in model population
Summary: What insights did we gain?
– Much of the complexity in developing a dynamic transmission model lies in the calibration
– A large-scale change in vaccination policy (e.g. targeted universal transition) makes a great test case
– A dynamic model is more challenging to work with than a static model, but can also give us deeper insights
– Our result: A Canada-wide switch from TIV to QIV is projected to be highly cost-effective across all plausible inputs– Province-specific analyses (AB, MB, ON, QC, NS) yield very
similar CE results
Backup slides
TIV and transmission dynamics: An interesting insight…– The WHO’s choice of B lineage to include in TIV matches the dominant
circulating B lineage in only ~50% of seasons
– Insight from 4Strain: The WHO actually does much better than this.
– …Why? Because circulation of TIV-included B lineage preferentially suppressed, which in many seasons actually changes the dominant lineage!
TIV actually works better than we think!
Even with perfect prediction, TIV would not prevent as many
outcomes as QIV
OR
Modeling the impact of a Canada-wide switch from TIV to QIV
comparator intervention difference % differenceage mean 95% CI - L 95% CI - U mean 95% CI - L 95% CI - U mean 95% CI - L 95% CI - U mean 95% CI - L 95% CI - U
influenza cases ALL 2,933,460 2,532,276 3,351,695 2,797,922 2,392,853 3,199,681 -135,538 -228,154 -76,677 -4.6% -7.7% -2.7%0-4 266,218 235,144 302,789 252,960 223,195 287,226 -13,258 -20,646 -8,264 -5.0% -7.6% -3.2%5-19 566,688 489,747 645,471 542,466 465,387 618,874 -24,221 -38,946 -14,626 -4.3% -6.8% -2.6%20-49 1,316,489 1,136,295 1,503,404 1,263,216 1,081,597 1,444,672 -53,273 -89,103 -30,831 -4.0% -6.7% -2.4%50-64 432,127 368,561 499,095 412,697 347,580 477,183 -19,430 -33,835 -9,993 -4.5% -7.8% -2.3%65-74 190,464 162,214 220,556 177,776 150,287 206,369 -12,688 -22,542 -6,051 -6.6% -11.7% -3.2%75-84 114,966 97,973 133,823 105,945 89,322 123,201 -9,021 -16,197 -4,119 -7.8% -13.7% -3.6%85-99 46,508 39,434 53,898 42,861 36,056 49,865 -3,647 -6,533 -1,629 -7.8% -13.5% -3.6%
GP visits ALL 1,066,568 921,034 1,218,892 1,014,368 868,298 1,160,118 -52,200 -88,460 -29,055 -4.9% -8.2% -2.8%0-4 121,129 106,990 137,769 115,097 101,554 130,688 -6,032 -9,394 -3,760 -5.0% -7.6% -3.2%5-19 179,920 155,495 204,930 172,229 147,764 196,482 -7,691 -12,366 -4,645 -4.3% -6.8% -2.6%20-49 412,061 355,660 470,566 395,387 338,540 452,182 -16,674 -27,889 -9,650 -4.0% -6.7% -2.4%50-64 135,256 115,360 156,217 129,174 108,793 149,358 -6,082 -10,590 -3,128 -4.5% -7.8% -2.3%65-74 118,088 100,572 136,745 110,221 93,178 127,949 -7,866 -13,976 -3,751 -6.6% -11.7% -3.2%75-84 71,279 60,743 82,970 65,686 55,380 76,385 -5,593 -10,042 -2,553 -7.8% -13.7% -3.6%85-99 28,835 24,449 33,417 26,574 22,355 30,916 -2,261 -4,050 -1,010 -7.8% -13.5% -3.6%
ER visits ALL 59,704 51,257 68,574 56,309 47,987 64,721 -3,395 -5,907 -1,731 -5.7% -9.7% -3.0%0-4 6,794 6,001 7,727 6,456 5,696 7,330 -338 -527 -211 -5.0% -7.6% -3.2%5-19 988 848 1,131 948 806 1,087 -41 -68 -24 -4.1% -6.8% -2.4%20-49 10,008 8,638 11,429 9,603 8,222 10,982 -405 -677 -234 -4.0% -6.7% -2.4%50-64 15,095 12,875 17,435 14,417 12,142 16,669 -679 -1,182 -349 -4.5% -7.8% -2.3%65-74 14,514 12,361 16,807 13,547 11,452 15,726 -967 -1,718 -461 -6.6% -11.7% -3.2%75-84 8,761 7,466 10,197 8,073 6,806 9,388 -687 -1,234 -314 -7.8% -13.7% -3.6%85-99 3,544 3,005 4,107 3,266 2,747 3,800 -278 -498 -124 -7.8% -13.5% -3.6%
hospitalizations ALL 32,986 28,319 37,886 31,110 26,512 35,757 -1,876 -3,264 -956 -5.7% -9.7% -3.0%0-4 3,754 3,316 4,269 3,567 3,147 4,050 -187 -291 -117 -5.0% -7.6% -3.2%5-19 546 469 625 523 445 601 -23 -37 -13 -4.1% -6.8% -2.4%20-49 5,529 4,772 6,314 5,306 4,543 6,068 -224 -374 -129 -4.0% -6.7% -2.4%50-64 8,340 7,113 9,633 7,965 6,708 9,210 -375 -653 -193 -4.5% -7.8% -2.3%65-74 8,019 6,829 9,285 7,484 6,327 8,688 -534 -949 -255 -6.6% -11.7% -3.2%75-84 4,840 4,125 5,634 4,460 3,760 5,187 -380 -682 -173 -7.8% -13.7% -3.6%85-99 1,958 1,660 2,269 1,804 1,518 2,099 -154 -275 -69 -7.8% -13.5% -3.6%
deaths ALL 4,836 4,114 5,606 4,508 3,811 5,230 -328 -584 -156 -6.8% -11.9% -3.2%0-4 11 9 12 10 9 11 -1 -1 0 -5.0% -7.6% -3.2%5-19 10 9 12 10 8 11 0 -1 0 -4.1% -6.8% -2.4%20-49 118 102 135 114 97 130 -5 -8 -3 -4.0% -6.7% -2.4%50-64 579 494 669 553 466 639 -26 -45 -13 -4.5% -7.8% -2.3%65-74 2,228 1,898 2,581 2,080 1,758 2,415 -148 -264 -71 -6.6% -11.7% -3.2%75-84 1,345 1,146 1,566 1,240 1,045 1,441 -106 -190 -48 -7.8% -13.7% -3.6%85-99 544 461 631 501 422 583 -43 -76 -19 -7.8% -13.5% -3.6%
comparator intervention differencemean 95% CI - L 95% CI - U mean 95% CI - L 95% CI - U mean 95% CI - L 95% CI - U
Cost of vaccination: $112,089,969 $111,646,794 $112,605,730 $151,441,924 $150,843,161 $152,138,755 $39,351,954 $39,196,367 $39,533,025Cost of GP visits: $45,169,166 $39,005,771 $51,620,087 $42,958,488 $36,772,434 $49,130,978 -$2,210,678 -$3,746,274 -$1,230,469Cost of ER visits: $13,217,880 $11,347,784 $15,181,519 $12,466,233 $10,623,828 $14,328,485 -$751,647 -$1,307,831 -$383,222Cost of hospitalizations: $114,493,950 $98,131,051 $131,727,254 $107,859,049 $91,782,364 $124,123,339 -$6,634,900 -$11,578,189 -$3,344,257Total payer costs: $284,970,966 $260,842,138 $310,472,595 $314,725,695 $290,749,551 $338,668,374 $29,754,729 $22,687,791 $34,327,516QALYs lost: 68,980 59,036 79,436 64,930 55,206 74,837 -4,050 -7,076 -2,033LYs lost: 45,675 38,909 52,852 42,732 36,152 49,573 -2,944 -5,215 -1,417
Cost of vaccination: $851,459,123 $848,060,111 $855,366,082 $1,150,384,894 $1,145,792,576 $1,155,663,489 $298,925,772 $297,732,465 $300,297,407Cost of GP visits: $344,113,857 $295,469,942 $394,623,287 $326,401,246 $278,359,229 $374,775,557 -$17,712,612 -$29,155,352 -$9,782,336Cost of ER visits: $100,348,798 $85,768,367 $115,510,962 $94,421,737 $80,104,557 $108,840,130 -$5,927,061 -$10,030,340 -$2,998,031Cost of hospitalizations: $868,685,572 $741,808,220 $1,000,472,625 $816,473,915 $691,417,165 $942,481,425 -$52,211,657 -$88,525,569 -$26,104,120Total payer costs: $2,164,607,350 $1,974,928,547 $2,361,559,430 $2,387,681,792 $2,203,932,794 $2,577,321,287 $223,074,442 $171,667,499 $259,231,329QALYs lost: 522,596 446,330 601,554 490,805 414,820 567,537 -31,791 -54,079 -15,845LYs lost: 344,912 293,245 398,169 322,013 270,885 374,069 -22,899 -39,878 -10,871
mean 95% CI - L 95% CI - UCost per case averted: $227 $97 $421Cost per GP visit averted: $596 $250 $1,126Cost per ER visit averted: $9,520 $3,792 $19,199Cost per hospitalization averted: $17,231 $6,863 $34,751Cost per death averted: $102,420 $38,591 $218,186Cost per QALY gained: $8,057 $3,175 $16,417Cost per LY gained: $11,344 $4,311 $23,953
Parameter fitting
Overall approach (analogous to that of van der Velde et al. 2007 for an HPV model):
– Prior ranges chosen for input parameters to be varied– Allowable target ranges chosen for model outputs – Sets of input parameters drawn using Latin hypercube sampling– One simulation run for each parameter set– Posterior parameter distribution consists of all parameter sets which
produce simulation outputs satisfying all the target ranges
Above approach used to fit natural history parameters of the model. Fitting targets are:
– “natural attack rate”, i.e. force of infection in the unvaccinated population, (Turner et al. 2003 HTA, using placebo arms of vaccine/antiviral RCTs)
– relative fraction of influenza A and B, by season (CDC surveillance data)– % of circulating influenza B covered by the B strain selected for vaccine
(Reed et al. 2012)
Can then perform simulations in different settings (i.e. with different demographics, vaccine uptake, etc.), each time drawing parameter sets from the above posterior distribution
Background: Ontario’s adoption of a universal influenza immunization program (UIIP)
Implemented in 2000; world’s first large-scale universal influenza immunization program
Resulting changes in both vaccine uptake and influenza-associated events have been studied in detail (Kwong et al., PLoS Medicine 2008). Events considered:– doctor’s office (GP) visits – emergency room (ER) visits– hospitalizations– deaths
Objective: Assess how well our model agrees with Kwong et al.’s results
Simulating Ontario’s universal influenza immunization program (UIIP): Model inputs I Population, birth, death rates from Statistics Canada, http://www5.statcan.gc.ca/cansim/
Simulated period is 1997-2004, as in Kwong (2008) (i.e. 3 yrs pre-introduction, 4 yrs post-introduction of universal influenza immunization
Uptake rates:– age 6-23 months: Campitelli et al. (2012)– age 2-11 years: extrapolated using Moran et al. (2009)– age 12 yrs and up: Kwong et al. (2008)
“natural attack rate”, i.e. force of infection in the unvaccinated population, (Turner et al. (2003) HTA, using placebo arms of vaccine/antiviral RCTs)
Simulating Ontario’s universal influenza immunization program (UIIP): Model inputs I Population, birth, death rates from Statistics Canada, http://www5.statcan.gc.ca/cansim/
Simulated period is 1997-2004, as in Kwong (2008) (i.e. 3 yrs pre-introduction, 4 yrs post-introduction of universal influenza immunization
Uptake rates:– age 6-23 months: Campitelli et al. (2012)– age 2-11 years: extrapolated using Moran et al. (2009)– age 12 yrs and up: Kwong et al. (2008)
“natural attack rate”, i.e. force of infection in the unvaccinated population, (Turner et al. (2003) HTA, using placebo arms of vaccine/antiviral RCTs)
fraction of circulating influenza B, and fraction of B covered by vaccine: FluWatch surveillance network
vaccine efficacy: Tricco et al. (submitted), systematic review
against influenza A
against influenza B, lineage match
against influenza B, lineage mismatch
outcomes probabilities:Pr(GP visit|flu), Pr(hospitalization|flu), Pr(death|flu): Molinari et al. (2007)
Pr(ER visit|flu): extrapolated from Kwong et al. (2008)
Simulating Ontario’s universal influenza immunization program (UIIP): Model inputs II