Assistant Professors Create Interactive Model that Makes COVID-19 Predictions Model featured on CDC Website for partner decision making
authored by Trista Sobeck
Imagine if a decision-maker could look into the future and see the impact of their choices. There would be less opportunity for mistakes. Time and even perhaps, lives, could be saved.
UC San Diego Assistant Professors, Yian Ma, of Halıcıoğlu Data Science Institute (HDSI), and Rose Yu, Department of Computer Science and Engineering and affiliated faculty at HDSI, have created a tool that can do just that. In a world that is currently experiencing the biggest pandemic in more than 100 years and has no cure, the need for data-driven decisions is of paramount importance.
With more than 50.6 million people across the globe who have COVID-19 and 1.26 million dead, optimizing real-time decisions is indeed necessary.
Both Ma and Yu began working at UC San Diego around the same time and immediately recognized the potential for interdisciplinary, multi-institutional collaboration. With Ma, a specialist in machine learning and focusing on theory and Bayesian inference, and Yu, an expert in deep learning and spatiotemporal modeling, a natural partnership in time series modeling began.
After they received professional encouragement from Alexandro Vespignani, Director and Sternberg Family Distinguished Professor at Northwestern University, a physics-deep learning hybrid forecasting model for COVID-19 began. Work on DeepGLEAM had started.
What is DeepGLEAM?
This digital interactive model is now displayed in real-time and allows users to see mathematical modeling at work. Living on the CDC Forecasting webpage, it helps the CDC and others respond to the COVID-19 pandemic. Decisions such as planning, resource allocation, social distancing measures, as well as other protocols can be increased or altered depending upon virtual outcomes.
Although the idea of a physics-based or data-driven model alone that can help predict outcomes and guide decisions is not new, UC San Diego’s approach is quite unique.
DeepGLEAM combines both physics and data and uses the signal of a discrete stochastic epidemic computational model GLEAM (Global Epidemic and Mobility Model) to inform a deep learning spatiotemporal forecasting framework and improve predictions to study the spatiotemporal COVID-19 spread.
More plainly, a spatiotemporal element allows for forecasting in a dynamic environment. It is used in applications from autonomous vehicles operations, to energy and smart grid optimization, to logistics and supply chain management. Capturing the element of mobility was key for DeepGLEAM. Perhaps even more plainly–people are mobile, so data must take physical movement into consideration.
Making Decisions in the public and personal arenas
Public health officials will be making decisions about when and where to allocate vaccines when they arrive. In addition, officials constantly need to know how many ventilators and hospital beds are needed and where. They must continually make informed decisions about where the next “spike” will take place and how many more resources should travel to a new location.
In order to follow through on these decisions, they need to know beforehand how many deaths, hospitalizations, or infections a certain county will be experiencing in the coming weeks.
Another audience that cannot be overlooked is a typical person concerned about COVID-19 who wants–or needs–to travel throughout the United States.
Americans make more than 405 million long-distance business trips per year, accounting for 16% of all long-distance travel, according to a preliminary analysis of the National Household Travel Survey (NHTS).
Because this movement is pertinent for a sustained business, mobility is a necessary piece of information to have at hand. DeepGLEAM is a hybrid model and uses real-world data about COVID-19.
It relies on data such as:
- when a person has been infected;
- where they have traveled;
- death records;
- travel, and
- re-opening restrictions from all 50 states.
To Teach and Impact Social Issues
Both Ma and Yu have even higher hopes that go beyond the execution of tools. They aspire to advance machine learning for social impact. Their research focuses on real-world problems and solves social challenges. From a technical perspective, their approach highlights the benefits of combining physics-based first-principle modeling with data-driven machine learning to tackle large uncertainty and limited observations.
What better dreams could one have than to touch the future, teach it to others, and predict what can happen in the future? It’s real. At Halıcıoğlu Data Science Institute at UC San Diego, it’s happening.
For questions regarding this article and other HDSI information, please contact HDSIComm@ucsd.edu.
Prediction has been traditionally approached via a model-based paradigm, i.e., (a) fit a model to the data at hand, and (b) use the fitted model in order to extrapolate/predict future data. Due to both mathematical and computational constraints, 20th century statistical practice focused mostly on parametric models. Fortunately, with the advent of widely accessible powerful computing in the late 1970s, computer-intensive methods such as the bootstrap and cross-validation have freed practitioners from the limitations of parametric models, and paved the way towards the ‘big data’ era of the 21st century. Nonetheless, there is a further step one may take, namely going beyond even nonparametric models. The Model-Free Prediction Principle is based on the simple notion of transforming a complex dataset to one that is easier to work with, e.g., i.i.d. or Gaussian. As such, it restores the emphasis on observable quantities, i.e., current and future data, as opposed to unobservable model parameters and estimates thereof. Coupled with resampling, the Model-Free Prediction Principle allows us to go beyond point prediction in order to construct frequentist prediction intervals without resort to unrealistic assumptions such as normality or linearity.
