In 2013, Dr. Tavpritesh Sethi, Physician-scientist and Assistant Professor of Computational Biology at Indraprastha Institute of Information Technology Delhi (IIIT-Delhi) led a study looking at the prevalence of diseases in India. On average, about 200,000 Chronic Obstructive Pulmonary Disease (COPD) patients would walk into the clinics on a single day all across India.

“We wanted to see what patterns could be learnt from these data obtained from vast different geographical locations within the same country. We took into considerations of all these contexts and something very interesting came up,” says Dr. Sethi in a recent AIMed interview. “We found that areas with low COPD are also areas with access to clean cooking fuels such as refined petroleum gas or electricity. If we are able to increase the penetration of clean cooking fuels, the prevalence of COPD will be cut by almost 50%”. That was when Dr. Sethi realized the power of context in artificial intelligence (AI).

Since then, Dr. Sethi and his research team began to develop their own software and making them available to the public. At the same time, they also combined the technology with deep learning and other AI techniques to standardize the field of contextual AI. “I came from a medical background; I moved to data science early on after gaining my degree in medicine. So, in my mind, unless I put context into place, I don’t think an AI model will be greatly useful. I guess it’s my clinical training and prior knowledge that is driving me”.

The meaning of context

According to Dr. Sethi, at the moment, most algorithms are trained to answer narrow questions with particular outcomes, such as whether the lung nodule is cancerous or not. For contextual AI, machines will be trained to learn the representations of the World. It is learning to see how different demographics, ethnicity to geographic, social mobility and socio-economic factors interact to derive at the given set of data.

“One example is our paper published in The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19). We used contextual AI, vision-decision network sort of learning approaches to have an insight into what would mitigate health inequality in the US? Also, how does the context influence how well a county perform in terms of healthcare indices, behaviors and so on. We put all the contextual factors into a model while not limiting it to predict one narrow outcome. Instead, it will learn the representations of the data and relationships between them,” Dr. Sethi explains.

“We then make an ensemble of these models because unlike narrow focused models which we put all the compute optimization into solving one problem. In representation learning problem, the model tries to learn the structures of the data and solve everything so things may go wrong if it’s not robust enough. At the end of the day, we are trying to create a mental makeup of the World, healthcare setting so to say and from there, address narrow AI problems that can help with predictions”.

Emphasis on human-machine collaboration

Dr. Sethi said one of the keys to contextual AI is human-machine collaboration. He thought AI fails to make healthcare more connected is probably because the lack of trust between clinicians and developers. He believes it is crucial for AI models assisting clinicians in decision making to consider the different contextual scenarios.

For example, a patient who walks into the clinic by him/herself is likely to be different from a patient who has been found on the street. Likewise, some patients may come from an area with a higher prevalence of diseases while others from places more prone to chronic conditions. These contexts, will influence diagnoses and subsequent interventions and recommendations. Most of the time, clinicians will make a mental note of the patients’ background and context before settling on the priorities and ailments.

“Machine is likely to fall through the cracks if they make errors that human clinicians won’t make and aren’t able to provide help in this very first step. So, it will become more and more important to have that context being given the same regard as AI continues to evolve. This can only happen for a strong collaboration between the algorithms and experts. This means contextual AI is highly human-centric.”