MEPI Subgroup Contributed Talks

Woldegebriel Assefa Woldegerima

York University

"The Mathematics of Deep Neural Networks with Application in Predicting the Spread of Avian Influenza Through Disease-Informed Neural Networks (DINNs)"

Deep learning has emerged in many fields in recent times where neural networks are used to learn and understand data. This study combines deep learning frameworks with epidemiological models and is aimed specifically at the creation and testing of DINNs with a view to modeling the infection dynamics of epidemics. Our research thus trains the DINN on synthetic data derived from an SI-SIR model designed for Avian influenza and shows the model’s accuracy in predicting extinction and persistence conditions. In the method, a twelve hidden layer model was constructed with sixty-four neurons per layer and ReLU activation function was used. The network is trained to predict the time evolution of five state variables for birds and humans over 50,000 epochs. The overall loss minimized to 0.000006, characterized by a combination of data and physics losses, enabling the DINN to follow the differential equations describing the disease progression.

---

Jongmin Lee

Department of Mathematics, Konkuk University

"How to Deal with the Health-economy Dilemma during a Pandemic: Research Framework and User-interactive Dashboard"

During the early stages of the COVID-19 pandemic, it was important to minimize both medical and economic costs. In this study, we introduce a machine learning-based multi-objective optimization framework that can propose cost-effective social distancing strategies. Our approach finds Pareto solutions that balance different goals, like reducing infections and minimizing social distancing costs. Then, the cost-benefit analysis can adjust each cost factor—for example, value of statistical life (VSL), fatality rate, or GDP. We also provide an interactive web dashboard so that policymakers and the public can test various scenarios easily. We tested this framework on the COVID-19 pandemic in Korea. The results show that the difference between the cost-optimal strategy and implemented strategy is 10% in cost. Notably, our results reveal two distinct patterns in cost-optimal solutions. When social distancing cost is proportional to intervention intensity, an on-off lockdown strategy proves most economical. In contrast, when the cost increases sharply with intensifying social distancing, maintaining a consistently moderate level of intervention minimizes overall expenses. By letting people explore different cost settings and intervention strategies, this tool can support more balanced decisions during emerging infectious disease crises in the future.

---

Asa Rishel

University of Maryland, College Park

"Mind over matter: balancing the benefits of COVID lockdowns with their cost on mental health"

The COVID-19 pandemic took its toll not only on the physical health of those who lived through it, but also on their mental health. I will present a model of the direct and indirect effects of COVID-19 and the associated public policies on mental health. This is an SIRS model of COVID-19, with compartments for mild, acute, and chronic COVID-19 infections and additional compartments for populations with mental health symptoms. Parameters are determined based on fitting from the first wave of COVID-19 in the New York state population, which includes several changes in local government policy, e.g, lockdown orders, which have an effect on the rate at which mental health systems develop. Finally, an additional “delay” term is included in the model to account for the delay between lockdowns going into effect and individuals developing mental health symptoms. The goal of our analysis is to understand how government policy in response to a pandemic can seek to maximize the population's quality-adjusted life years (QALY), which is a measure not only of lifespan, but also the quality of the years lived. I will present some preliminary results suggesting the optimal timing and strength of government lockdown mandates.

---

Arsene Brice zotsa ngoufack

Université du Québec à Montréal

"Stochastic epidemic model with memory on the previous infection and with varying infectivity and waning immunity"

After an individual has been infected by a pathogen, T lymphocytes store information about the pathogen. Consequently, upon reinfection by the same pathogen, an immune response memory is triggered. This immune memory allows the body to react very quickly against the pathogen. Indeed, when an individual recovers from a virus, sometimes the individual acquires full immunity. In some cases, the individual's immunity persists for some period, after which it decreases progressively and can even disappear. I will then present a stochastic epidemic model with memory on the previous infection, incorporating varying infectivity and waning immunity. More precisely, we will present a functional law of large numbers when the size of the population tend to infinity. We will also present results on the behaviour of the epidemic, more precisely the threshold for the existence of an endemic equilibrium, and study the stability of the endemic equilibrium.

---

Phoebe Asplin

University of Warwick

"Estimating the strength of symptom propagation from synthetic data"

Symptom propagation occurs when an individual’s symptom severity is correlated with the symptom severity of the individual who infected them. Determining whether - and to what extent - these correlations exist requires data-driven methods. In this study, we use synthetic data to determine the types of data required to estimate the strength of symptom propagation and investigate the effect of reporting bias on these estimates. We found that even a relatively small number of contact tracing data points was sufficient to gain a reasonable estimate for the strength of symptom propagation. Increasing the number of contact tracing data points further improved our estimates. In contrast, population incidence alone was insufficient to accurately estimate the symptom propagation parameters, even with a large number of data points. Nonetheless, concurrently using population incidence data with contact tracing data led to increased accuracy when estimating the overall disease severity. We then considered the effect of severe cases being more likely to be reported in the contact tracing data. When contact tracing data alone was used, we found that our estimates for the strength of symptom propagation were robust to all reporting bias scenarios considered. However, the reporting bias led us to overestimate the overall disease severity. Using population incidence data in addition to contact tracing data reduced the error in disease severity but at the cost of increasing the error in the strength of symptom propagation when reporting bias was in both primary and secondary cases. Consequently, these errors led to us sometimes finding support for symptom propagation, even when the synthetic data was generated without.

---

Emma Fairbanks

University of warwick

"Semi-field versus experimental hut trials: Comparing methods for novel insecticide-treated net evaluation for malaria control"

We aim to compare results for the predicted reduction in vectorial capacity caused by pyrethroid and pyrethroid-piperonyl butoxide insecticide treated nets (ITNs) between semi-field Ifakara Ambiant Chamber tests (I-ACT) and experimental hut experiments. Mathematical modelling and Bayesian inference frameworks estimated ITN effects on mosquito behavioural endpoints (repelled, killed before/after feeding) to predict reductions in Anopheles gambiae’s vectorial capacity for Plasmodium falciparum transmission. The reduction in biting estimates are generally greater for I-ACT, possibly due to lower mosquito aggression: Although I-ACT vectors are probing before release, experimental hut vectors are actively seeking a blood meal. I-ACT estimates higher probability of killing vectors which have fed, while experimental huts show greater killing before feeding, possibly due to their open-system design, where vectors can contact the net, then attempt to exit and get trapped. This is supported by most of the mosquitoes being caught before feeding being in the exit trap. While the I-ACT is a closed system, were vectors cannot exit or be trapped, increasing the likelihood of returning to host-seeking and feeding. Despite these differences, both methods yielded similar predictions for the overall reduction in vectorial capacity. Results suggest that I-ACT provides a good initial assessment of the impact of adulticide modes of action of these nets. Challenges of semi-field experiments include how to model the change in efficacy from practical use over time. However, important advantages include the ability to easily trial different strains of vector (including different resistance levels) and allowing rapid data collection. Parameterising models with location-specific bionomic parameters allows for setting -specific predictions of the impact of different nets, with the potential to include additional modes of action for other active ingredients.

---