PROMPTS: How an AI tool is educating pregnant women in Kenya through SMS

As artificial intelligence (AI) research advances and innovators come up with new ways of deploying the technology in various fields to augment the human brain, one of the challenges has been going past accessibility, high costs and infrastructure issues to boost reach.

AI-powered SMS chatbots are coming in handy, especially among Kenyans in the low-income bracket, by delivering large language models (LLMs) to low-cost feature phones, locally referred to as ‘kabambes’, via SMS.

These adaptable virtual assistants learn and use natural language processing (NLP) to understand what human beings are asking via SMS, then use AI to respond in a human-like way. 

NLP is a branch of AI that enables computers to comprehend, generate, and manipulate human language.

One such tool is PROMPTS, an AI-enabled support tool providing personalized SMS advice in English and Swahili to new and expecting mothers across Kenya to tackle maternal deaths.

An expectant mother signs up for the platform, developed by the Kenyan non-profit Jacaranda Health, by sending an SMS.

They then answer questions about themselves sent via text, from which the model creates profiles of them to give free personalised AI-generated advice on any questions they ask about their pregnancy journey. Through the SMSs, the platform also sends the mothers appointment reminders.

The tool becomes a woman's virtual obstetrician on their feature phone, similar to how smartphone users use ChatGPT, the popular generative AI chatbot by American AI research company OpenAI.

In Kenya, the non-profit founded by Nick Pearson in 2016 has partnered with the Ministry of Health (MoH) and county governments to deploy PROMPTS in public health centres when mothers go for prenatal care.

They are also working with community health promoters to boost uptake in villages in some of the 23 counties they currently cover as they seek to cover all of Kenya, where 6,300 women die each year during pregnancy and childbirth in the country, per MoH figures.

For training – the process of feeding AI models with data sets to help them learn to make accurate responses to queries – the questions the mothers are asking PROMPTS has been central in refining the tool.

“We use the huge database of the questions these mothers are asking to train our AI model on what kind of tone to use when responding to them, empathy and even greetings,” Jay Patel, Jacaranda’s Director of Technology tells Citizen Digital.

Additionally, the organisation says it combines data from mothers on their pregnancy experience and their feedback into dashboards and shares with health facilities and system managers towards targeted improvements.

HUMAN HELP DESK

But then there is the accuracy concern, as with any AI product. While large language models can now generate human-like responses as research progresses, they are still hallucinating.

Hallucination is the generation of false or misleading responses presented as fact by AI systems, usually because they are based on flawed or insufficient training data.

Once a user sends in a question, the AI model classifies it on an ‘urgency’ scale – low on the scale are questions about things like what to eat at different pregnancy stages or if something is safe.

The AI model handles this and responds through a tailor-made SMS, meaning it takes into account the mother's age and pregnancy stage based on the information they gave upon sign-up.

The goal is to make it factor in, says Patel, things like the prevailing weather conditions in the mother’s region.

High-urgency questions on the other hand are about, say, a severe headache or vaginal bleeding.

In this case, the mother is referred to an officer at Jacaranda’s help desk where doctors engage with her one-on-one and help refer her to a health centre and, if need be, an ambulance from the organisation’s partner hospitals.

“If it is urgent, we do not want AI talking to them; it will write a response, but the response will go to the help desk, who then decide how to act on the matter by providing a response or calling the mother and refer her to the hospital,” Patel says.

“After she is treated, we then call two days later to follow up on her progress.”

Additionally, Patel says they have also adopted two AI tools; one for writing the answers and another to audit the responses.

“If a response passes the 90 per cent score, it is when we send it to the user,” he says.

Jacaranda says PROMPS has reached 3 million women so far across over 1,200 health centres in Kenya. According to Patel, they process over 10,000 questions daily.

The organisation is also in Eswatini and Ghana and they are currently testing the product in Nigeria and Nepal.

They are grant-funded and have also been sharing deployment costs with county governments to keep the expenses low.

“For us, end-to-end, it is about a dollar per user up to a year post-partum,” says Patel.

Among the organization’s backers is the US tech giant Google, which through its charity arm Google.org last year gave Jacaranda a $1.4 million grant and fellowship to advance the AI tool.

In March this year, Google picked Jacaranda among 21 non-profits for its first Generative AI Accelerator.

Now, as they work on prototyping voice support to make users able to speak to the tool, the organisation says it recently open-sourced the LLM powering PROMPTS.

This makes it customizable for anyone seeking to develop similar AI-driven support solutions in different languages, new channels like WhatsApp or voice, or new content like mental health or agriculture.