Schedule

Many modern (machine) learning strategies depend on the intelligent acquisition of informative samples. Such sampling methods can be viewed as an instantiation of the exploration-exploitation problem. Initially, one is unclear about the state of the environment and the goal is to take observations that refine the understanding of the state. If one has a series of “experiments” (or queries), each of which provide information about the state, an important question is how to design that sequence of experiments to enable a decision about the environmental state as quickly as possible. Exploration-exploitation problems abound in applications such as anomaly detection, target localization, dynamical system tracking, medical diagnosis, wireless body area sensor networks etc. The problem of experiment design for classification (hypothesis testing) has been persistently studied since the 1940s. Then and now, there has been an emphasis on the design of asymptotically optimal methods. Herein, we will provide new analysis which enables the design of strategies for the finite sample regime. In key cases, our methods are also asymptotically optimal, but provide significantly improved finite sample performance. We specialize our analysis to the problem of anomaly detection for which we can determine asymptotically tight upper and lower bounds on the misclassification error and provide an experiment design strategy with excellent finite sample performance. We further consider the application of our approach to group-testing, wherein different experiments call for the pooling of samples which can dramatically reduce the number of experiments needed. Finally, we consider the problem of testing of populations to provide good spatial estimates of the incidence of an anomaly, such as SARS-CoV-2 positivity. We have preliminary analysis of SARS-CoV-2 serological tests based on randomized testing undertaken by a colleague at USC’s School of Public Policy. Our proposed strategy suggests that uniform allocation for randomized testing over heterogeneous regions may not yield the best estimates of positivity rates and offers a method by which active hypothesis testing can be used to improve such estimates.

We study a spatial epidemic model, which explicitly accounts for population movements, and propose an optimization framework for obtaining targeted policies that restrict economic activity in different neighborhoods of a city at different levels. We focus on COVID-19 and calibrate our model using the mobile phone data that capture individuals’ movements within New York City (NYC). We compare the targeted policy with a uniform policy and demonstrate significant increases in allowed employment for the same reductions in infections as in a uniform policy. Our targeting framework gives policy makers an approach for curbing the spread of epidemics while limiting unemployment.

When the Covid-19 pandemic spread to central Europe, task forces consisting of experts in different fields were formed in most countries to offer advice to political decision-makers on various aspects of the epidemic. One work package of the task force in Luxembourg focused on modelling and monitoring the epidemic evolution. As part of this task force modelling group, we developed an agent-based model for simulating various scenarios of epidemic development and effects of different mitigation measures. The infectivity model is based on a multi-layered social network between agents, where different layers can be switched on and off, depending on the prevailing policy. A standout feature of our model is that household, workplace and school class information was obtained from the social security system of Luxembourg, and the whole population is included in the model with realistic demographic structure. The social network approach enables detailed simulations taking into account population heterogeneity while also keeping the computation time reasonable. This, in turn, allowed simulating a high number scenarios with multiple randomised replicates. The results were used to support political decision-making, particularly during the deconfinement phase during spring and summer of 2020.

The COVID-19 crisis highlighted the importance of non-medical interventions, such as testing and isolation of infected individuals, in the control of epidemics. Here, we show how to minimize testing needs while maintaining the number of infected individuals below a desired threshold. We find that the optimal policy is adaptive, with testing rates that depend on the epidemic state. Additionally, we show that such epidemic state is difficult to infer with molecular tests alone, which are highly sensitive but have a short detectability window. Instead, we propose the use of baseline serology testing, which is less sensitive but detects past infections, for the purpose of state estimation. Validation of such combined testing approach with a stochastic model of epidemics shows significant cost savings compared to non-adaptive testing strategies that are the current standard for COVID-19.

A month after COVID-19 arrived in Uruguay in March 2020, the President decided to appoint an Honorary Advisory Group of Scientists (GACH). As one of the three coordinators of this group, it was my task to organize a variety of efforts in quantitative monitoring and mathematical modeling of the progress of the pandemic, and distill recommendations. The job entailed unprecedented public exposure and, for six months, collected praise as Uruguay largely avoided the first wave with relatively few restrictions through social distancing and Testing, Tracing and Isolation (TeTrIs). The following six months, however, were challenging as the next wave exceeded the Tetris capabilities, and eventually built to high levels in April-May 2021, with severe impact on the local health system. Control was restored by the month of July, after an aggressive vaccination program which has reached substantial coverage. In this talk I will reflect on the lessons of this experience, the contributions and limits of modeling, and the interplay between science and politics.

The COVID Pandemic was an unprecedented challenge for universities across the world. How could operations continue in person without unacceptable risk of life and illness? Purdue decided to pursue methods of safe in-person education shortly after lockdowns began in March. The ability to open safely was achieved in large part through data-driven decision making. Executives and the medical team received up-to-the-minute analytics built off of a sophisticated, Purdue-developed, data ecosystem and digital contact tracing system. These tools worked together to contain the spread of COVID and allow Purdue to continue operations in person from Fall 2020 onwards.

The COVID-19 pandemic that has been sweeping the world since early 2020 is distinct from earlier pandemics in one important respect. In earlier communicable diseases, persons who were incubating the virus but did not fully develop the infection were not capable of infecting others. In contrast, in the case of COVID-19, a majority of those infected are "asymptomatic," and show no or very few symptoms. Yet they are capable of infecting others, and their rate of viral load shedding (both in quantity and time duration) is comparable to that of persons manifesting symptoms. Thus, in order to analyze the spatial and temporal spread of the COVID-19 pandemic, it is essential to be able to estimate the number of asymptomatic patients using only available measurements of positive test results, hospitalizations, discharges, and deaths. The present team of authors have presented a series of models that can be used to make accurate predictions about the pandemic, both in India and around the world. Their first model, published in the Indian Journal of Medical Research, built upon an earlier paper of Robinson and Stilianakis (2013), in which it is assumed that ALL asymptomatic patients go undetected. A later model called SUTRA (which is submitted to the 2021 CDC) takes cognizance of two important aspects. First, because of contact tracing, some fraction of asymptomatic patients ALSO get detected. Second, a pandemic "spreads" through a region, and it is therefore important to introduce and estimate the "reach" of the pandemic as a time-varying (and monotonically increasing) parameter. The SUTRA model has been successfully used to predict the course of the pandemic in India and in 21 other countries around the world.

In the paper propose for the workshop, the authors pursue the next logical steps in modelling the COVID-19 pandemic, by incorporating two important aspects. First, they include a loss of immunity on the part of recovered patients, using an exponential decay rate. Currently available estimates indicate that those who recover from COVID-19 retain their immunity with a mean time of around eight months. Second, they include the impact of vaccination in the midst of a pandemic. Traditional vaccination models in the literature are usually based on the assumption that the previous wave of the pandemic has receded, and the next one has not yet started. In contrast, in the case of COVID-19, countries around the world have begun to vaccinate their populations even as the pandemic continues. In this aspect, the authors include the "efficacy" of vaccination as a parameter, and make quantitative predictions about the impact of efficacy on the future course of the pandemic. Though the authors' motivation is to study the COVID-19 pandemic, their methods are widely applicable.

We describe agent-based models to predict the spread of COVID-19 at a large public university, similar to the University of Illinois at Urbana-Champaign. The purpose of the modeling was to determine whether or not it is possible to reopen such a university without exponential growth of the epidemic on campus, and what surveillance testing, and other non-pharmaceutical interventions are required to accomplish this goal. Distinctive features of this modeling approach are: (1) Physical modeling of transmission of SARS-CoV-2 by aerosols; (2) Representation of student behavioral heterogeneity through a range of infection zones, with differing transmission characteristics and risks; (3) Evaluation of infection through classroom exposure using the social network generated by the timetable of over 45,000 students; (4) Estimation of the effective reproduction number achievable by a series of non-pharmaceutical interventions. The models do not assume compliance with social distancing or other health authority guidelines, outside the university, but do assume (initially at least) compliance with testing and legally required isolation as instructed by health authorities. The reopening of the University of Illinois at Urbana-Champaign was accomplished over the late Summer and Fall semester 2020, with approximately a million RT-PCR tests delivered to the campus community, and with approximately 4000 positive cases identified, predominantly amongst undergraduates. Our work shows how a blend of physical, behavioral, and epidemic modeling can be used to predict trends and semi-quantitative features of COVID-19 spread in congregate settings such as a large public university.

Starting from October 2020, the cross-disciplinary project CRUSH Covid is run at Uppsala University, Sweden, to support the authorities of the Uppsala Region in decision-making regarding resource management to mitigate the effects of the pandemic. Mathematical modeling and prediction of COVID-19 spread is part of this effort and the authors are in charge of it. The project conceived the concept and coined the notion of data-driven interventional epidemiology. Whereas traditional epidemiology concentrates on the causes of health outcomes and diseases in populations, the new discipline will rather focus on epidemiological situation awareness through data fusion and in-time appraisal of effective and sustainable interventions. The data-driven part is naturally involved with predicting the epidemiological situation on the basis of information obtained from relevant sources, e.g., the healthcare providers, through symptom monitoring applications, sewage water analysis, mobility monitoring, and other. Even perfect predictions would not change the course of infection spread but precisely informed epidemiological interventions could. Therefore, the prediction of the epidemiological state is communicated to the public health authorities in the Uppsala Region for shaping and implementing proper interventions. The efficacy of the interventions can be quantified and evaluated by calculating, e.g., the reproduction number from historical data. The talk will thus cover the general setup of model-based estimation and prediction of epidemiological state, the parameter and state estimation techniques that have been found useful, and concrete examples of informed interventions implemented in the Uppsala Region within the framework of CRUSH Covid.