It's been more than a year since all of us were locked indoors due to the ongoing?SARS CoV-2 pandemic that wreaked havoc on our lives.?
While most of us have now started to step out again and things are beginning to normalize, the pandemic is showing no signs of slowing.
But this entire pandemic has been quite a learning experience for us all -- doctors, businesses and even us regular individuals. And no one really wants to relive this pain once again. This is also enabling governments and institutions around the world to be prepared, in case the next pandemic comes knocking on our doors (instead of being utterly unprepared like the COVID-19 pandemic) and a duo of Indian-origin AI experts are trying to help bridge this gap.
Mudit Jain, an AI engineer at Google and Biplab Deka, PhD from University of Illinois, UC -- both IIT Kanpur graduates, teamed up to create an Artificial Intelligence-based solution for predicting COVID-19 cases accurately and prescribing intervention plan solutions.?
They understand that every location is different and will have different patterns. They¡¯ve come up with a scalable solution to this problem by creating separate AI models for each region to predict Covid-19 time-series in those regions.?
They built this solution on a traditional epidemiological technique called SEIR model and extended it with AI to make it a dynamic model instead of the traditional static approach.
Mudit explains "SEIR stands of Susceptible-Exposed-Infected-Recovered. It's a 90 year old modeling technique which models the number of people in each of these categories using differential equations and static assumptions.¡±
He added, ¡°We extended this SEIR model using AI and made the model dynamic i.e. location and time-dependent. So those static assumptions have been replaced by dynamic data-dependent outputs of AI models, thus making the whole model more accurate.¡±
For training the models, they made use of existing publicly available features like demographics, health infrastructure and economic indicators.
They also accessed Google Maps Mobility Index which is collected from Android phones worldwide in an anonymized and privacy secure way and which measures population movement changes in public areas such as parks, hospitals, workplaces, groceries, railways stations, airports etc.?
Another unique feature they used was Google's Covid-19 symptom search trends data which is similarly aggregated anonymously and contains the popularity of Google search queries related to Covid-19 symptoms.
Mudit shares how their AI-based implementation can offer a more accurate prediction of cases in future, "Using location wise (i.e. country or state wise) model predictions, healthcare policymakers can more accurately estimate the number of cases in coming months. That can help them proactively decide preventative measures to take such as travel restrictions, varying degrees of lockdowns, health awareness advertising campaigns etc.¡±
He added, ¡°They can also prioritize vaccination programs to target more vulnerable regions first. The model being dynamic, will also adapt to the impact of these measures and update predictions as new data arrives."
Mudit explained how pandemics, despite the difference in the perpetrator causing it, actually are not that different from one another in terms of outcome, "Mathematically all past, current and future pandemics follow a similar process and equations. This is why all public health experts were talking about "flattening the curve" and "R-nought values" even before Covid-19 started to spread widely.¡±
He added, ¡°Since the underlying mathematical modelling for each pandemic is the same, even though the virus's biological mechanism may be quite different, this AI-augmented SEIR technique can be used for future pandemic modelling too."
The AI tool built by Mudit Jain and Biplab Deka is one of the finalists for the Cognizant XPrize Pandemic Response Challenge, victory in which will help them bag a $500,000 cash prize. They¡¯ve managed to reach the finals based on the accuracy of their predictions, application of novel AI techniques and the ease of interpretability of their predictions.?