Agent-based network models(ABNM) are suitable for simulating individual-level dynamics of infectious diseases. However, as ABNM simulates a scaled-version of the full population, consisting of all infected and susceptible persons, they are computationally infeasible for studying certain questions in low prevalence, such as during new or re-emerging disease outbreaks, or slower spreading diseases such as HIV, Hepatistis C, or Tuberculosis. For e.g., simulating 100,000 nodes to represent the U.S. population would generate only 400 people with HIV. We developed an Evolving Contact Network Algorithm, combining network theories with machine learning, to simulate networks of only infected persons and immediate contacts and thus dynamically evolve the network as the infection spreads.
Molecular analyses, using nucleotide sequences of the virus from persons diagnosed with an infection, can identify clusters or groups of infections that are genetically very similar. Thus, they help detect disease outbreaks, measure growth, and rapidly respond if needed. However molecular analysis is limited to persons who are diagnosed and have a sequence, and thus, generate only partially observable sub-networks of full networks. We developed mathematical methods to generate clusters and the underlying full network, specifically, a cluster generation algorithm generates clusters replicative of those observed using molecular data and maps clusters to the full contact and transmisison network. Using molecular methods for early detection and the model for understanding true size of network could further help with rapid response and optimal response.
Typically, dynamic models of infectious diseases simulate behaviors and model transmissions and disease progression as functions of behaviors. However, social conditions are among key drivers of behaviors that increase risk of infection. We develop statistical machine learning methods to first model behaviors as functions of social conditions. The resulting model will help quantify disparities, determine social needs of affected populations, and evaluate equitable allocation of intervention resources.
Reinforcment learning algorithms are suitable for evaluation of optmial sequence of decisions for eliminating infectious disease epidemics. However, they are computationally burdensome, and when combined with computationally large simulation enviromnents, especially for noncovex problems, such as in infectious disease models, become impractical to use. Through suitable choice of proxy action and state space we signifcantly reduce the size of the problem and generate an action space that makes the problem convex.
Typically, in infectious disease analyses, we develop models specific to each disease and conduct intervention analyses related to that disease. However, social determinants are key drivers of infectious diseases, and thus, structural interventions are key for overall disease prevention. Therefore, joint modeling multiple related diseases, such as diseases with common modes of transmission, would help collectively evaluate strategies to strengthen overall public health efforts. However, given the vastly varying epidemiology of diseases, there is no one suitable simulation technique that is suitable. Agent-based network models are suitable for slower spreading diseases, whereas differntial equations based models are suitable for faster spreading diseases. We developed a new simulation technique that simulates persons with atleast one slow spreading disease in a network and all other persons including those with only fast spreading diseases in a compartmental model, with an evolving contact network algorithm maintaining the contact dyanmics between the two populations.
National and global public health strategic plans and guidelines inform allocation of public health resources to varying intervention programs. We develop national-level stochastic dynamic simulation models, for use by public health agencies, for identifying optimal intervention combinations. We have developed models for HIV, COVID, breast cancer, colorectal cancer, and cervical cancer.
In the context of infectious diseases, intervention decisions are typically made by multiple independent agents, e.g., state or national jurisdictions, cooperatively or non-cooperatively, but in an environment (diseases) that naturally involves cross-jurisdictional interactions. In an optimization context, infectious disease models typically do not consider these dynamics or simulate single jursidictions in isolation ignoring jurisdictional interactions, or conduct specific scenario based evaluations in a non-optimization context. Though recent advances in RL such as in multi-agent reinforcement learning would be suitable to address this gap, MARL algorithms are computationally impractical to implement in a national or global context as computational complexity grows exponentially with number of agents (jursidictions). We aim to develop suitable alternative algorithms.