Cost-effectiveness of maternal influenza immunization in Bamako, Mali: A decision analysis

Background Maternal influenza immunization has gained traction as a strategy to diminish maternal and neonatal mortality. However, efforts to vaccinate pregnant women against influenza in developing countries will require substantial investment. We present cost-effectiveness estimates of maternal influenza immunization based on clinical trial data from Bamako, Mali. Methods We parameterized a decision-tree model using prospectively collected trial data on influenza incidence, vaccine efficacy, and direct and indirect influenza-related healthcare expenditures. Since clinical trial participants likely had better access to care than the general Malian population, we also simulated scenarios with poor access to care, including decreased healthcare resource utilization and worse influenza-related outcomes. Results Under base-case assumptions, a maternal influenza immunization program in Mali would cost $857 (95% UI: $188-$2358) per disability-adjusted life year (DALY) saved. Adjusting for poor access to care yielded a cost-effectiveness ratio of $486 (95% UI: $105-$1425) per DALY saved. Cost-effectiveness ratios were most sensitive to changes in the cost of a maternal vaccination program and to the proportion of laboratory-confirmed influenza among infants warranting hospitalization. Mean cost-effectiveness estimates fell below Mali’s GDP per capita when the cost per pregnant woman vaccinated was $1.00 or less with no adjustment for access to care or $1.67 for those with poor access to care. Healthcare expenditures for lab-confirmed influenza were not significantly different than the cost of influenza-like illness. Conclusions Maternal influenza immunization in Mali would be cost-effective in most settings if vaccine can be obtained, managed, and administered for ô$1.00 per pregnant woman. We built a decision tree model of the costs and benefits of maternal influenza immunization. All benefits of maternal influenza vaccine were assumed to stem from prevention of laboratory-confirmed influenza in the pregnant mother, the infant, or the post-partum mother. After an initial decision to either vaccinate or not vaccinate the pregnant mother, further events including influenza infection in the pregnant woman, infant, or post-partum mother proceeded in a probabilistic manner (Fig 1). At each node of influenza infection, a sub-tree determined the associated monetary costs from treatment and the loss of disability-adjusted life years (DALYs) (Fig 2). Each infection was stratified by severity as requiring no treatment, outpatient therapy only, or inpatient therapy. Healthcare encounters including influenza requiring outpatient or inpatient therapy were each associated with monetary costs of illness. The outcomes of maternal death, stillbirth, and infant death each resulted in a loss of DALYs as a function of the life expectancy at the time of the event [15], with maternal death estimated at 24.7 years (the average age at enrollment) [14]. All branches were identical in the vaccine and no-vaccine branches except in the probability of contracting influenza. The incremental cost-effectiveness of maternal influenza immunization was calculated by dividing the difference in net costs in the vaccine and no-vaccine branches by the difference in DALYs lost between the two branches. Squares designate decision points. Circles designate probabilistic events. Triangles designate terminal nodes. Ellipses (…) indicate a symmetrical sub-tree that is not shown due to space constraints. (A) Series of events that determine the monetary and utility costs of influenza infection during pregnancy. (B) Series of events that determine the monetary and utility costs of influenza infection during the first 6 months of life. All calculations were based on a societal perspective. To reduce the likelihood of overestimating the cost-effectiveness of maternal influenza immunization, we only included DALY losses from premature death [15] and costs related to the acute episode; long-term sequelae were excluded from the model. Thus, we did not specify an analytical horizon, discount rate, or base year for dollars. Attack rates of influenza in mothers and infants, vaccine efficacy in mothers and infants, the risk of infant hospitalization after contracting LCI, and outpatient and inpatient costs of illness for mothers and infants were all taken from primary data collected during a phase 4 randomized controlled effectiveness trial conducted in Bamako, Mali from September 2011 –January 2014 that determined the efficacy, safety, and immunogenicity of trivalent inactive influenza vaccine (TIV) in pregnant women and their infants up to 6 months of age. Full study details are described elsewhere[14] but are summarized below with an emphasis on collection of cost data. Additional parameters for which the trial in Mali was underpowered were obtained through literature review (Table 1). Methods of estimation for all parameters are given in detail in the Supplementary Information (S1 Table). The threshold for cost-effectiveness was set to the 2013 gross domestic product (GDP) per capita in Mali, US$715 [14]. CFR = Case-fatality ratio; LBW = low birth weight; DALY = Disability-adjusted life year. Pregnant women were recruited during their 3rd trimester at 6 community and referral health centers in Bamako, Mali. After randomization, women were vaccinated with either TIV or meningococcal conjugate vaccine (MCV). From enrollment until the infant reached age 6 months, field personnel performed weekly visits, during which the participating woman and infant (if already born) had their temperatures measured and were evaluated for ILI. When the ILI case definition was met nasopharyngeal/oropharyngeal swabs and a malaria blood smear were obtained, and a team dedicated to estimating costs of illness was contacted. For participants treated as outpatients, study personnel visited the home every 2–5 days for the duration of the episode to ascertain the direct costs of illness (including medications, labs, traditional healers, and transportation) and indirect costs of illness (defined as the number of workdays lost by each family member multiplied by that family member’s average daily earnings). Additionally, all medications prescribed by study physicians and filled at the recruitment health centers were reimbursed by the study and counted as direct costs. Infants with severe ILI warranting hospitalization were admitted to l’Hôpital Gabriel Touré, a public university hospital, and were visited by a study physician daily who accounted for direct and indirect costs. In addition to direct costs to families that were reimbursed through the maternal influenza study, other diagnostic tests (complete blood count, blood cultures, chest radiographs) were provided free of charge to some patients through separate ongoing studies when indicated; the costs of these tests were added back into total cost estimates. Inpatient episodes included all direct and indirect costs incurred in both the outpatient and inpatient setting for a single ILI episode and were modelled with a log-normal distribution. Outpatient episodes included illnesses with non-zero total costs but with no hospitalization; these were also modelled with a log-normal distribution. All episodes included zero and non-zero costs of outpatient and inpatient episodes and were modelled with an exponential distribution. Statistics were performed using Matlab Version 7.7 (The Mathworks, Inc). Approval for the research was obtained from the University of Maryland, Baltimore Institutional Review Board; the ethics committee of the Faculté de Médecine, Pharmacie et Odonto-Stomatologie of Mali; and the Ministry of Health of Mali. We performed univariate and multivariate sensitivity analyses to ascertain the major determinants of the cost-effectiveness of maternal influenza immunization. First, each input variable was individually changed by a relative proportion to assess its impact on the CER. Second, we performed a regression tree analysis on 10,000 Monte Carlo simulations that randomly sampled all input variables across their distributions simultaneously. The regression tree algorithm found the threshold value of the single input variable that best divides all simulations into two groups, minimizing the variance in CERs within each group. This process yielded the single parameter and threshold value with the highest predictive power in estimating the CER. The algorithm then performed the same calculation on each respective branch to determine the next most important parameters in predicting the CER and repeated until the number of simulations in each leaf was <10. Finally, the tree was pruned to prevent over-fitting using a 10-fold cross-validation strategy that minimizes generalization error. Under clinical trial conditions, active surveillance likely led to earlier detection of illness, higher healthcare resource utilization, and better health outcomes. For example, low birth weight incidence in the Mali trial was significantly lower than that found in Bamako in Demographic and Health Surveys (DHS) [14,16]. Thus, we modeled different levels of access to medical care (Fig 3) by varying the proportion of ill individuals who received care at the level warranted by disease severity (inpatient, outpatient, or no care required). The health outcomes and costs were adjusted according to whether the individual received the appropriate level of care; a person who did not receive adequate care had a higher likelihood of death but did not incur the costs of outpatient or inpatient illness. (A) Changes to the sub-tree for influenza infection during pregnancy. (B) Changes to the sub-tree for influenza infection during the first 6 months of life. Finally, we performed a two-way sensitivity analysis to look at the impact of changing attack rates on vaccination strategies. Season-to-season and geographical variability in influenza incidence may profoundly affect cost-effectiveness. Additionally, vaccinating earlier in pregnancy might provide greater protection throughout the pregnancy, but may result in less antibody transfer and protection of the infant. Conversely, vaccinating in the 3rd trimester likely maximizes infant protection, but leaves the pregnant woman unimmunized for a larger proportion of the pregnancy. We therefore examined how the interaction of these two attack rates affected the CER. The model was built and analyzed using the Treeplan, Sensit, and Risksim macros (Decision Support Services, San Francisco) in MS Excel as well as Matlab Version 7.7 (The Mathworks, Inc). We calculated 95% confidence intervals (CI) for primary cost data and 95% uncertainty intervals (UI) for Monte Carlo simulations.