Models emulating the spread of infectious diseases help design optimal reactive measures and prepare for pandemics
Industry Example: Health Care, Government
Who Cares: Decision makers, Health Care, Health Care Research, Financing
What Enabled: Optimal policy development, National-scale infectious disease models of real-time efficiency, Interactive What-if scenarios
Collaborators: Lander Willem and Prof. Philippe Beutels (Centre for Health Economics Research & Modeling Infectious Diseases (CHERMID), University of Antwerp, Belgium), Sean Styven and Prof. Jan Broeckhove (Dept. Computer Science, Antwerp University, Belgium), Prof. Niel Hens ((Centre for Health Economics Research & Modeling Infectious Diseases, University of Antwerp and Center for Statistics, Hasselt University Belgium)
Models of spread of infectious diseases are important tools to inform policy makers about the risks of pandemics and anticipate optimal intervention measures. State of the art is in the large-scale simulators modeling the spread of infection using millions of individual-based models. While the most realistic, these simulations are the most computationally intensive (with every scenario taking several hours to compute). They have too many knobs to turn (controllable input variables), are too complex to understand fully and too ``black-box'' to discover best and worst-case scenarios to react and prevent the pandemics.
Data-driven modeling methods like DataModeler's symbolic regression are ideally suited to create efficient meta-models exactly mimicking the behavior of the behemoth computationally expensive simulators. From a variety of benefits two are critical - meta-models give instantaneous predictions of simulator outputs and therefore allow running interactive real-time what-if scenarios, meta-models only contain variables that impact the outcomes and therefore create focus for decision makers, meta-models are ensembles of explicit mathematical equations, which can be fully optimized, for example to identify optimal reactive measures for each type of disease.
In this large-scale interdisciplinary project with many collaborators we used the expertise in infectious disease modeling, parallel computing, iterative data-driven system understanding with symbolic regression, optimal design of simulation experiments with maximal information content and interactive visualizations to understand and exploit individual-based simulator models.
We used a publicly available state-of-the-art stochastic individual-based simulator for influenza epidemics written in C++, called FluTE (Chao D, Halloran M, Obenchain V, Longini I (2010) FluTE, a publicly available stochastic inuenza epidemic simulation model. PLoS Comput Biol 6: e1000656. doi: 10.1371/journal.pcbi.1000656 ). It simulates a population with realistic social contact networks and transmission probabilities based on the natural history of influenza.
Design of experiments helped us optimally run individual-based simulations for urban populations of up to 15 million of people using minimal numbers of experiments while getting maximal amount of information. Data-driven system understanding allowed to iteratively use all collected data to focus on the discovered drivers - reproduction number R0 (a critical parameter of a disease - the number of people an infected person can infect per day in a fully susceptible population), number of infected individuals seeded into the population, vaccination coverage and vaccine efficacy. The target outcome parameters were total clinical attack rate (the total number of infected people in the period of 160 days and nights), and the day of the peak of pandemic.
The existence of reliable and efficient meta-model estimating the probability of the pandemics spreading and effectiveness of reactive measures is a bliss in threat of a real pandemics. The parameter defining the infectiousness of a disease can be estimated in a few days, however these estimations are not precise and the real reproduction numbers can vary. After the reproduction rate is estimated from real data - different scenarios of similar ranges can be run in a matter of hours. On top of that, the unexpected worst-case scenarios can be identified automatically from the models, and analyzed. This not only gives the decision maker the power of knowing what will happen and what to do, but also quantifies the consequences of not taking the measures.
The interactive tool to run what-if scenarios for varying inputs is available online at: http://www.idm.ua.ac.be/
Related Links/ Publications/ Resources:
Lander Willem, Sean Stijven, Ekaterina Vladislavleva, Jan Broeckhove, Philippe Beutels, Niel Hens – Active Learning to Understand Infectious Disease Models and Improve Policy Making PLOS Computational Biology, April 2014, DOI: 10.1371/journal.pcbi.1003563
Key Words: policy, data-driven decision making, epidemiological models, vaccination, DataModeler Pro, what-if scenarios, simulation-based optimization