Using machine learning to predict short-term disease progression
University of Waterloo
Their
model, the Sparsity and Delay Embedding-based Forecasting model, or SPADE4,
uses machine learning to predict the progression of an epidemic using only
limited infection data. SPADE4 was tested on both simulated epidemics and real
data from the fifth wave of the Covid-19 pandemic in Canada and successfully
predicted the epidemics' progressions with 95 per cent confidence.
"Covid taught us that we really need to come up with methods that can predict with the least amount of information," said applied mathematics PhD candidate Esha Saha, the lead author of the study. "If we have a new virus emerge and testing has just started, we have to know what to do in the short-term."
When
a disease breakout occurs -- whether for new infections like Covid-19 or
existing ones like Ebola -- being able to predict the development of the
disease is essential for making public policy decisions.
"That's
what policymakers need right at the beginning," Saha said. "What
should we do in the next seven days? How should I allocate resources?"
Traditionally,
epidemiologists prefer to build and use complex models to understand the
progression of epidemics. These models, however, have several drawbacks, Saha
said.
They
require complex demographic information that is frequently unavailable at the
beginning of an outbreak. Even if that detailed information is available, the
models may not accurately reflect the complexity of the population or dynamics
of the disease.
The
Waterloo research team's new model addresses these drawbacks.
"By the time we're working on vaccines and cures, we're looking at longer-term data," Saha said. "But when a new disease arrives, this method can help give us insight into how to behave."