EmpatIA pilot report - our findings in enhancing healthcare for remote areas in Peru
A blog by Alexis de Brouchoven, a Frontier Tech Coach
Pilot: EmpatIA - AI to enhance healthcare in remote areas of Peru
Beginning in September 2022, the EmpatIA pilot aimed to improve healthcare access for remote communities in Peru. Through deploying a mobile AI chatbot, Avatr, real time monitoring and reminders could improve the healthcare experience and widen access across different communities.
Through a series of six sprints, the EmpatIA team worked to answer a variety of questions during the pilot phase, including validating user needs, finding hospitals and healthcare providers to deploy the tech, and designing and implementing a patient study.
The six core findings during the pilot phase are summarised below.
Validating user needs
Survey results indicated that 89% of patients found Avatr helpful for medication reminders, especially outside Lima, while private healthcare providers mentioned potential cost-saving and efficiency benefits. Specifically on regions out of Lima, we learned that some healthcare issues could be remotely addressed from Lima.
Although a partnership with EsSalud (one of the main healthcare providers in Peru) did not materialize, their insights helped identify physical mobility as a key access barrier. The technology showed indicative compatibility with public healthcare, and post-pilot interest from private providers indicated market demand.
Use of Avatr could increase access to healthcare
The pilot led a study to find whether the Avatr AI chatbot could help improve healthcare access in Peru as an intervention. The results showed that while it helped with medication adherence and provided valuable patient well-being data through patient output, there were areas for improvement, such as managing post-surgery pain and anxiety, and ensuring patients could effectively use the app. The small scale nature of the study limited the statistical significance, but highlighted the potential of digital solutions to bridge healthcare access gaps.
Demonstration effect of private sector innovation is key to public sector uptake
Although risk aversion and political instability hindered the opportunities for the pilot to deploy in the public sector via EsSalud, there was a strong indication that the agency was keen to explore the incorporation of AI within its healthcare service delivery. The Ministry of Health, whom was engaged in a limited capacity, signalled its interest in how this solution could eventually be part of its healthcare programmes.
One key learning was that the private sector emerged as a more viable pathway for introducing new technologies, with streamlined processes and a greater willingness to invest in innovation. Partnerships with private institutions like Detecta highlighted the appetite for launching initiatives in digital healthcare solutions.
It is possible to adapt Avatr to new contexts and specific users’ needs
A core goal of the pilot was to adapt Avatr for the Peruvian context, which entailed testing multi-lingual capabilities, multi-modal content, and user-friendly features tailored to patients' needs through customisations. Customisations included tracking pain levels for physician monitoring, providing educational content via videos, offering dietary guidance, and ensuring communication in Spanish. Multi-modal capabilities were tested by integrating relevant images into responses, though challenges with image extraction and accuracy were identified for future improvement. Overall, these modifications served to demonstrate Avatr’s suitability for specific users’ needs.
Avatr can work in multiple languages, however existing LLMs in Quechua and Aymara need further refinement before Avatr can incorporate them
The pilot explored adapting Avatr to support Peru's linguistic diversity, aiming to include Peruvian Spanish and indigenous languages like Quechua and Aymara, which are crucial for reaching remote areas. While Peruvian Spanish integration was successful, Quechua and Aymara language models performed poorly in conversational response tests due to insufficient training data. Although it was not feasible to refine these models during the pilot, with greater training data there is potential to enhance the tool even further.
Read the pilot report in full by clicking below:
If you’d like to dig in further…
🚀 Explore this pilot’s profile page
📚 Read the ecosystem report
📝 Check out the pilot’s kick-off post
