IndiaHealthcare and life sciences2022 – ongoing

ARMMAN

Machine learning to decide who gets live health-worker calls in the mMitra program

A Google Research model cut engagement drops by about 30% in a 23,000-woman field trial of an Indian maternal health program

WorkingIndependent studyVendor ↗

Problem

ARMMAN's mMitra program sends automated voice messages to pregnant women and new mothers, but many stop listening. The nonprofit has few health workers, so only a small number of women can get a live call.

Approach

Google Research and ARMMAN built a model that picks which women should receive the limited calls, using a restless multi-armed bandit method. For new enrollees with no listening history, it groups women by past listening patterns and maps demographics to those groups. About 23,000 women were split into three groups: standard care, round-robin calls and model-chosen calls. The study ran for seven weeks.

Results

23,003participants in the seven-week field studyIndependent study[1]

~30%reduction in engagement drops in the model-selected group versus comparison groupsIndependent study[1]

32%reduction in cumulative engagement drops versus standard care after seven weeks (p = 0.044), per Google ResearchVendor's case study[2]

Takeaway

The model allocated limited staff calls and reduced disengagement from health messages by about 30%. The results covered one seven-week study.

Sources

Field Study in Deploying Restless Multi-Armed Bandits: Assisting Non-profits in Improving Maternal and Child Health
Proceedings of the AAAI Conference on Artificial Intelligence28 Jun 2022Research paper or report
Using ML to Boost Engagement with a Maternal and Child Health Program in India
Google Research24 Aug 2022Vendor publication

Source summaries

[1]Proceedings of the AAAI Conference on Artificial Intelligence28 Jun 2022

Researchers from Google Research, Harvard, ARMMAN and Singapore Management University describe a bandit-based system for allocating limited service calls. In a seven-week study of 23,003 participants, the model group showed a statistically significant improvement over comparison groups, cutting engagement drops by about 30%. The researchers describe it as the first field demonstration of this method in public health.

[2]Google Research24 Aug 2022

Google Research explains that mMitra sends preventive-care voice messages from pregnancy to a child's first birthday and describes the three-group design. It reports a 32% cut in cumulative engagement drops versus standard care after seven weeks, while round-robin calls cut drops by 5.2%, which was not significant. The study measured only short-term results with limited call capacity.

Full details are in the original sources.

Added 25 Oct 2026 · Updated 25 Oct 2026