Finding optimal vaccination strategies for pandemic influenza using genetic algorithms

https://doi.org/10.1016/j.jtbi.2004.11.032Get rights and content

Abstract

In the event of pandemic influenza, only limited supplies of vaccine may be available. We use stochastic epidemic simulations, genetic algorithms (GA), and random mutation hill climbing (RMHC) to find optimal vaccine distributions to minimize the number of illnesses or deaths in the population, given limited quantities of vaccine. Due to the non-linearity, complexity and stochasticity of the epidemic process, it is not possible to solve for optimal vaccine distributions mathematically. However, we use GA and RMHC to find near optimal vaccine distributions. We model an influenza pandemic that has age-specific illness attack rates similar to the Asian pandemic in 1957–1958 caused by influenza A(H2N2), as well as a distribution similar to the Hong Kong pandemic in 1968–1969 caused by influenza A(H3N2). We find the optimal vaccine distributions given that the number of doses is limited over the range of 10–90% of the population. While GA and RMHC work well in finding optimal vaccine distributions, GA is significantly more efficient than RMHC. We show that the optimal vaccine distribution found by GA and RMHC is up to 84% more effective than random mass vaccination in the mid range of vaccine availability. GA is generalizable to the optimization of stochastic model parameters for other infectious diseases and population structures.

Introduction

Influenza is a major public health concern. Influenza spreads rapidly in seasonal epidemics which cost society a considerable amount in terms of health care expenses, lost productivity, and loss of life. Globally, influenza annually costs from $71 and $167 billion and results in 250 000 and 500 000 deaths (World Health Organization, 2004). Annual influenza epidemics occur partially due to strains of influenza genetically drifting from year to year. The influenza vaccine produced before the influenza season is targeted against the strains that are predicted to circulate in the coming season. However, a major antigenic shift can occur with little warning, resulting in pandemic influenza (Kilbourne, 1975). The two most recent pandemics were the “Asian Flu” A (H2N2) of 1957–1958 (Elveback et al., 1976, Jordan, 1961, Longini et al., 1978) and “Hong Kong Flu” A (H3N2) of 1968–1969 (Davis et al., 1970, Elveback et al., 1976, Longini et al., 1978, Sharrar, 1969). The Asian Flu is estimated to have caused 70 000 deaths in the US, while the Hong Kong Flu is estimated to have caused 34 000 deaths in the US. In addition, the Hong Kong pandemic of 1968–1969 is estimated to have cost society 3.8 billion dollars in the US alone (Kavet, 1972). Should a major antigenic shift occur, there may be time to produce only a limited amount of vaccine efficacious against the new strain and ensuing pandemic. Knowledge on how to distribute the limited supply of vaccine optimally among different age groups would help us to minimize the impact of the epidemic. This impact can be measured in many ways, two of which are number of illnesses and loss of life.

Optimization methods have been developed for deterministic simulation models (Anderson and May, 1991, Greenhalgh, 1986, Hethcote and Waltman, 1973, Wickwire and Guest, 1976), and optimization studies have been done to find the vaccine distributions for influenza using a simple deterministic model (Longini et al., 1978). However, they have not been done with more complex stochastic simulation models. Influenza is transmitted in a complex way from person to person. In addition, given an introduction of influenza into a population, the probability of a major epidemic and the possible size of an epidemic are highly variable. Thus, the mathematical models for influenza epidemics should have a detailed contact structure and be stochastic. In addition, the epidemic process is non-linear since the incidence of new infections depends on the current number of both infectives and susceptibles in the population at a particular time. All of these factors make optimization based on traditional gradient methods, such as the Newton–Raphson method, difficult or even prohibitive. Robbins and Munro (1951) developed a stochastic approximation method whose convergence is guaranteed under mild conditions. The method, however, requires knowledge of the analytic gradient of the considered objective function (Weisstein, 2004). Kiefer and Wolfowitz (1952) developed an extension to the Robbins–Munro algorithm. However, in terms of simulation optimization, the drawback to both of these methods remains the unavailability of gradients. In the case of our stochastic simulation multi-dimensional optimization problem, we consider genetic algorithms (GA) and random mutation hill climbing (RMHC) as stochastic optimization methods.

In this paper, we find optimal distributions of a limited supply of vaccine in the event of pandemic influenza generated by a stochastic simulation model using GA (Holland, 1975) as well as RMHC (Forrest and Mitchell, 1993). We configure the model to simulate the baseline illness attack rates consistent with the past patterns of Asian and Hong Kong pandemic influenza. We find optimal vaccine distributions in terms of minimizing influenza illness or death. These optimization methods are generally applicable to other infectious diseases and population structures.

Section snippets

The simulation model

We use a discrete time, stochastic simulation model of influenza spread within a structured population of 10 000 individuals to estimate the effectiveness of various distributions of a given amount of vaccine. This model is a direct extension of an earlier model developed for influenza intervention studies (Halloran et al., 2002a, Longini et al., 2004). The model simulates the stochastic spread of influenza in a population where the age and family structure approximate information from the US

Initialization

  • 1.

    Randomly generate an initial set of 50 individuals. We generate the initial pool of individuals such that (3) is randomly determined for each individual under the condition that (2) is satisfied at equality.

Iteration

  • 2.

    Evaluate each of the 50 individuals in the pool by pre-vaccinating the proportion of the population according to the individual being evaluated, then running the stochastic simulator 20 times. The resulting fitness of each individual is the mean ofi=15σiniwiover the 20 runs of the simulator. A smaller value of this fitness function represents a more fit individual, as this fitness function is essentially a loss function.

  • 3.

    Sort the 50 individuals according to their fitness.

  • 4.

    We select the best 25

Convergence

  • 6.

    We repeat this process from Step 2 to Step 5 until the best individual does not change for four consecutive generations. The best individual in the final generation will yield the optimal vaccine distribution given the quantity of available influenza vaccine, the particular population structure, the illness attack rate pattern, and the objective of the optimization.

Initialization

  • 1.

    Generate a random individual vA as defined in the GA approach which satisfies (2) at equality.

  • 2.

    Evaluate vA by pre-vaccinating the proportion of the population according to vA running the stochastic simulator 20 times. The resulting fitness of vA is the mean ofi=15σiniwiover the 20 runs of the simulator.

Iteration

  • 3.

    Randomly mutate vA by first creating a copy of vA, vB(1), and mutating vB(1) in the following manner:

    (a) Select one age group, i, at random, and set vBi(1)=vAi+U,where U is randomly chosen from a Uniform(0,1-vAi) distribution. A Gaussian distribution is commonly used to mutate continuous genes, however we noticed no difference in the quality of resulting individuals and running time of the algorithm using the uniform distribution. The advantage of using a uniform distribution in this case is

Convergence

  • 6.

    We continue this process until 100 consecutive mutations of vA do not reveal a more fit individual than vA.

Optimal vaccine distributions

Table 2, Table 3, Table 4, Table 5 show the results given by the GA. The results from RMHC are similar to those from the GA, and are thus not in the tables. For example, the RMHC result for the Asian-like flu case when minimizing illness attack rate and V/n=0.40 is v=(0.88,1.00,0.21,0.26,0.00). We do not expect exact results from each algorithm as the results are approximations of the optimal distributions. The two algorithms converging to significantly different distributions of the vaccine

Discussion

The optimal vaccine distribution is sensitive to the nature of the spread of the influenza agent, the objective for control, and also the amount of vaccine available. GA, as well as RMHC, in conjunction with a properly calibrated simulation model would allow us to obtain an approximation of the optimal vaccine distribution given that we understand the behavior of the next pandemic agent and that we determine an objective for the control of the agent.

It is of interest to compare the optimal

Acknowledgment

This work was partially supported by grant R01-AI32042 from the National Institute of Allergy and Infectious Diseases.

References (35)

  • R.B. Belshe et al.

    Correlates of immune protection induced by live attenuated, cold-adapted, trivalent, intranasal influenza virus vaccine in children

    J. Infect. Dis.

    (2000)
  • Census Bureau, United States, 2001. Census 2000;...
  • L.E. Davis et al.

    Hong Kong influenzathe epidemiologic features of a high school family study analyzed and compared with a similar study during the 1957 Asian influenza epidemic

    Amer. J. Epidemiol.

    (1970)
  • L.R. Elveback et al.

    An influenza simulation model for immunization studies

    Amer. J. Epidemiol.

    (1976)
  • D. Fedson

    Pandemic influenza and the global vaccine supply

    Clin. Infect. Dis.

    (2003)
  • Halloran, M.E., Longini, I.M., Struchiner, C.J., 1999. Design and interpretation of vaccine field studies. In: Monto,...
  • M.E. Halloran et al.

    Community trials of vaccination and the epidemic prevention potential

    Vaccine

    (2002)
  • Cited by (150)

    • Foresighted medical resources allocation during an epidemic outbreak

      2022, Transportation Research Part E: Logistics and Transportation Review
    View all citing articles on Scopus
    View full text