Researchers from Microsoft, the Indian Institute of Technology (IIT) and Tata Consultancy Services (TCS) Research have released a preprint paper that describes an Artificial Intelligence (AI) algorithm which helps cities and regions make policy decisions about lockdowns, closures, and physical distancing in response to pandemics such as COVD-19. 

While the paper is yet to go through peer-review, the coauthors claim that the AI algorithm is superior to the other available modelling tools because it learns policies automatically as a function of disease parameters. Disease parameters are indicators such as infectiousness, gestation period, duration of symptoms, probability of death, population density, and movement propensity. 

If the alogirthm bears out the researchers’ claims, the framework could be useful to organisations and governments in the nearly 200 countries with cases of the coronavirus. Asian nations including Singapore and Taiwan have demonstrated that containment strategies like contact tracing — the process of identifying people who may have come into contact with an infected person — can effectively mitigate the spread of disease.

The coauthors first generated a graph network containing objects that correspond to vertices; each pair of vertices is called an edge — with 100 nodes and 1,000 individuals. Each node represented a city or a region containing a certain number of individuals, and the strength of the connections between pairs of nodes was directly proportional to the product of the population between nodes and inversely proportional to the square root of the distance between them.

The researchers assumed that an open node allowed people to travel to and from other open nodes throughout the study. In addition, they also assumed that people showing symptoms weren’t allowed to travel to other nodes but asymptomatic and exposed people could do so and the fact that while symptomatic people were quarantined within nodes, a small number of people broke quarantine and circulated within. They also established baseline lockdown policies for the algorithm in which they assumed that each node had the option to be locked down or opened once per week.

For the next step, the researchers fed the AI algorithm the best disease parameters implemented for COVID-19 such as an incubation period of 5-10 days, an infected period of 7-14 days, an 80% likelihood of showing visible symptoms, a 2% death rate, and a 100% transmission probability for infected persons who come into contact with susceptible persons. The algorithm runs multiple stimulations to generate reliable statistics. They then defined a set of policies that locked down any given node if the fraction of symptomatic people in that node crossed a predefined threshold of 5%, 10%, 20%, 50%, or over 100%. 

With all these functions and parameters in place, the team deployed a Deep Q Network reinforcement learning (RL) algorithm. An RL rewards the algorithm for every correct decision whilst punishing it for the wrong choices. Over time, the algorithm made a per-node binary decision each week — “open” or “lockdown” — by running a number of simulations of the spread of the disease. The researchers quantified the cost of two scenarios - “open” or “lockdown”. Each day of lockdown and each person infected was given the weightage of 1.0 and each death was given a weightage of 2.5. The algorithm was rewarded for outcomes with lower costs. 

They ran 75 stimulations over the course of 52 weeks for experimentation. The researchers determined that policies with 5% to 10% lockdowns experienced a lower peak of infections. Predictably, the policy was wary of decisions contributing to an increase in the fraction of symptomatic people within the same node and the population overall, and so it locked down larger nodes earlier once the infection started spreading and nodes where the potential for outside infection was higher as soon as infection began spreading within the node.

The coauthors caution that none of the authors are experts on communicable diseases and that the AI model in the study doesn’t account for population size and geography, and that they didn’t use real data for the network model. But they say that a deeper analysis is in progress and that they’ll continue to add more detailed descriptions and literature review in stages.