What can AI do? This is a question that many companies, organizations, and even individuals are trying to answer. If you ask Google Deepmind, the company’s AI division, or the NGO Armman, their answer would be: AI can save the lives of women in India. For the past several months, Deepmind, in collaboration with Armman, has been using its uniquely named AI Restless Bandit to ensure that pregnant mothers stay connected to prenatal care, thereby reducing dropout rates and improving health outcomes.

“We want to benefit people who have not benefited from AI,” says Dr. Milind Tambe, Principal Scientist and Director of AI for Social Good at Google DeepMind.

While AI advancement is often associated with tools like ChatGPT and Gemini that write code and novels, Google DeepMind is working on an AI model with a very cool name – Restless Bandit AI – to reduce maternal mortality in India, starting with Mumbai.

According to the 2018-2020 report of the Sample Registration System (SRS), India’s maternal mortality ratio (MMR) is 97 per 100,000 live births. Excessive bleeding during or after delivery is a major cause of maternal mortality in India.

In 2019, Tambe met Dr. Aparna Hegde, founder of an NGO called Armman, at a Starbucks in Mumbai. Hegde’s NGO has long been using technology to improve health outcomes for underprivileged pregnant women and children, but there was a gap they needed help filling. The organisation has a call centre, where health workers field a large number of calls from pregnant mothers.

The NGO was educating over 350,000 underprivileged mothers in Mumbai about pregnancy, nutrition and prenatal care, and sending 2-minute automated messages from the time a mother conceived to the time she gave birth. However, Armman faced a significant dropout rate of 30 to 40 percent. To help fill that gap, Google’s Restless Multi-Arm Bandit was born. It was a design to “increase the efficiency of (the NGO’s) calling process.”

Fun fact: The name of the AI ​​model is inspired by the imagery of a gambler standing in a row of slot machines, often called a ‘one-armed bandit’. The gambler has to figure out which machine he has to play, how many times he has to play them, in what order, and whether he has to stay with the current machine or switch to another one. Hence, the name: Restless Multi-Armed Bandit.

Tambe says the AI ​​model identifies who may be at risk of dropping out and how to best allocate calling resources to achieve the most effective outcome. Throughout the pregnancy and delivery journey, Armaman workers send around 140 automated messages to pregnant mothers. Restless Bandit’s task is to determine which mothers health workers should call for personalized follow-up, who needs the most support, who is more vulnerable, the frequency of contact, to ensure that most of these mothers stay connected for as long as possible during their pregnancy.

“This model predicts the changing behavior of mothers over time,” says Tambe. “It identifies mothers who would benefit from this model.”

The director of AI for Social Good explains that out of a large list of mothers who need to be contacted, Restless Bandit AI can predict based on demographic information which mother is most likely to respond to these calls and benefit from them. This prediction is based on their past behaviour, previous pregnancies, etc. In fact, Tambe mentions that Armaman’s staff noticed that the AI ​​model “could somehow identify mothers who are facing any problem or difficulty in their pregnancy”. With its prediction accuracy, the model helped reduce the dropout rate by 30 percent. The company says that only listener data and demographic data of mothers were used for these predictions.

With Restless Bandit AI, Google was able to help arms workers reach 30 percent more mothers, getting their families involved in their health routines, taking better supplements, getting them better educated about maternal health so they can avoid complications at birth.

Along with Armman, Google is also working on a national programme called Kilkari, under which the company will work with 3 million mothers across the country. Tambe says the Restless Bandit AI model can also be used in TB prevention. TB patients often need to take six months of medication, making it easy for them to drop out of the programme. Using the model, health workers will be able to identify patients who would benefit the most from personalised interventions or visits by health workers to remind them to take their medication.

Tambe says the key to collaboration between tech companies, academic institutions and nonprofits in tackling public health issues is to “start with the problem first.” In its collaboration with Armman, Google Deepmind began by analyzing the problem and then developing AI models and solutions to address it, he says. The approach should be problem-focused rather than trying to forcibly fit models or solutions into issues, and that’s the advice Tambe gives to other organizations that want to leverage AI for social good, especially in the health sector.