The CSIR-IIITH-Intel Covid Project has been commissioned to deploy faster & better methods for Covid-19 testing as well as develop risk stratification algorithms. It is helmed by Prof U. Deva Priyakumar and Prof Vinod PK of the International Institute of Information Technology, Hyderabad. A true example of government-academia-industry partnership, the project is carried out in collaboration with the Council of Scientific & Industrial Research (CSIR), Government of India, and Intel India.

How it started

“It is around March when Intel approached us to see if we could collect data and then develop some models that can be used for risk stratification and mortality prediction. Originally we started with risk stratification only,” reveals Prof Deva in a conversation with INDIAai.

Risk stratification is a technique for systematically categorising patients based on their health status and other factors. It estimates the probability of a person succumbing to a disease or benefiting from a treatment for that disease.

“Not only this, Intel also was talking to Institute of Genomics and Integrative Biology (CSIR-IGIB). The idea is to see if we can sequence these samples from Covid positive patients, and then see how the variations in the genome will affect the risk.”

“One is based on the clinical parameters that we collect from the hospitals. And then the second one is through genomic information. That's how this whole thing started,” says Prof Deva about the two ongoing initiatives.

“Currently we are writing a paper using the genomic information and then trying to see what kind of mutations in the virus are possibly responsible for higher severity or high risk," he adds.

Unpuzzling the puzzle

Explaining how risk stratification worked, Prof Deva explains, “The idea was simple. Even before we started working with Max Hospital Delhi, we already had quite some data from other countries which showed that mortality and the number of people being affected is a very, very less. At that point, people were still trying to understand what were the factors which influenced the severity of the disease. The one thing that we understood was that of all the people who were getting affected, the younger people were recovering faster.” 

So age was a clear factor that determined survival.

The role of IIIT-H

“Our contribution is in getting the data from Max Hospital and CSIR-IGIB and then developing the machine learning models.”

Elaborating on the data collection method, Prof Deva says, “Mortality rates are very different in different countries, right? For example, at that point, we realised that the mortality rate in India was far less compared to the West. So now we also wanted to see if there are any variations within the country. While the Max hospital data came from Delhi, we started working parallelly with Gandhi hospital in Hyderabad. We also made initial efforts to get something from Kolkata and Rajasthan, but as you know getting health data is extremely difficult”

Prof Vinod PK of IIIT-D (left) in conversation with Jibu Elias of INDIAai (right)

Data collection for Indian patients

“It took some time before we could get our hands on any kind of data pertaining to Indian patients. This is because there are procedures in healthcare data such as getting the ethics clearance.”

However, the professors decided to proceed with whatever data was available. “We found this work based on data from some hospital in Wuhan that was collected in the January-February period of 2020. That's the data set which had mortality data. There was reasonable distribution of people who survived and people who did not survive, and hence we developed that model first based on Wuhan population. And finally, we started working on the Indian data much later, around August,” says Prof Deva.

What are the inputs for the AI model?

"All the data collected was only from patients who were RT-PCR positive and admitted to the hospital. And the hematological parameters, like the blood work and then the vitals and then the need for ventilation or need for supplemental oxygen all the way up to the outcome. There could be one of three possible outcomes: patient is admitted in hospital but without much complication; patient had to be given supplemental oxygen or they had to be put on ventilator; or patient does not survive."

The team finally shortlisted five features: age, neutrophils, lymphocytes, LDH, and hs-CRP.

How accurate are the AI models?

Prof Vinod highlights a pertinent point. “The idea is that we want to predict the risk or mortality well in advance, not just three or four days before the outcome when it's already too late. The whole idea of predicting this is to give extra care to those patients, given that the healthcare system is over-burdened. Therefore, we can actually improve the efficiency of the healthcare system."

Prof Deva explains, "The model accuracy or performance depends upon the current data set that we have right now. So if you look at the risk stratification models, the accuracy is more than 70%. And for the mortality predictions it is more than 60%, keeping in mind this is the data set from only five 544 patients from Max hospital, Delhi. So its only one geographical location and there are only 11% mortality cases out of the total 544. Since there is imbalance of data within that cohort, so you can expect the performance to be at that level, which is more than 70%. But then if you compare the performance with the Wuhan data set, it reaches more than 90% because there is very good balance of data: around 50% of the cases there are mortality cases."

"We believe that this performance that we have got right now can improve and that more and more data that has been collected"

Adding to this, Prof Vinod says, "When we say 70% accuracy, what we mean is the prediction about 12 days before the outcome. That's the main advantage that we are able to predict the risk or mortality very well in advance when it is really not possible for a human expert to say anything. This way, patients can be better cared for."

Speaking of practical application of this model, Prof Deva says, "So if you want to take this model into a clinical setting, there is always a chance that we need to prove that this works on different cohorts from different parts of India. And that it is also comparable with the predictions of the doctor who's diagnosing the patient. So all these things have to be done before we could translate them into clinical practice."

Can the model be applied during India's second wave?

Since the model was built on the data from the first wave, there are some refinements to be done before the model can be applied during the second wave. "It is very different now because a considerable section of the population has been vaccinated," says Prof Vinod. He adds, “We started with the Wuhan data even though we knew that it's not going to be useful for our cohort and our population, but then again, we started so that as soon as the relevant data comes, we want to just get going.”

Of course, it helps to begin early and plan ahead of time. “Now we have all resources: the algorithms are ready and the codes are written. As soon as we get the latest data, we will rapidly be able to come up with very good models for this wave as well. And that will make a difference.”

Major challenges along the way

Access to good quality data remains the biggest roadblock for AI researchers. 

Prof Vinod says, “Getting an electronic health record of the patient is still not easy in the Indian setting compared to Western countries. Therefore, getting to a point where we systematically collect those data and then make it accessible after anonymisation is still the bottleneck that we feel we have to break if we want to create impact with AI or healthcare informatics."

But that’s not the where the problems end for the academicians; any research work is truly fruitful only when it come to life. “We are happy to come up with a model and possibly publish. But then there is also a step further from there where one has to translate that or take it to the clinical practice where we have to work with startups or any other players who are interested in seeing the end outcome of this project,” says Prof Vinod.

Image by Tumisu from Pixabay