More mothers keep listening to health messages

The mMitra service-call scheduler. [1]

A restless multi-armed bandit policy used Whittle indices. K-means grouped past listening patterns, and a random forest assigned new participants to groups. [1]

The 2021 paper identifies the algorithms and training approach, but not a numbered software release. [1]

Researchers from Google Research, Harvard, and Singapore Management University worked with ARMMAN. [1]

The model ranked participants for staff calls within the weekly calling limit. [1]

The system modeled two weekly listening states and used 40 groups derived from historical data. [1]

Registration details and listening logs produced a weekly list of participants for staff to call. [1]

The model selected call recipients. ARMMAN staff placed the calls and supported participants. [1]

Participants received recorded health messages. Each week, the scheduler selected a subset for live calls that encouraged continued listening. [1]

ARMMAN staff made the calls without knowing the participants’ assigned study group. [1]

Machine learning estimated listening behavior, and a bandit policy allocated staff calls. [1]

The scheduler selected who received a call. Comparing it with round-robin calling tested that selection policy. [1]

Not assessed

The scheduling policy reduced loss of engagement with health messages. [1]

Continued access to maternal and child health information. [1]

Observed in a randomized field study. [1]

Mothers enrolled in mMitra and ARMMAN’s outreach program. [1]

Pregnant women and new mothers enrolled in the program. [1]

The study included 23,003 women. The model group contained 7,668, of whom about 20% received a staff call. [1]

Limited staff time meant that only some participants received calls. The measured benefit was an aggregate change in listening. [1]

Listening was measured over seven weeks. [1]

Sustained listening was the intended benefit. [1]

The program could call only a small share of participants each week. The study did not establish which clinical outcomes followed. [1]

AIBD-0099-E01, AIBD-0099-E02, AIBD-0099-E03 [1] [4] [5]

Researchers randomly assigned participants to model-selected calls, round-robin calls, or regular messaging alone. [1]

Call logs identified weeks in which a participant listened to at least one message for more than 30 seconds. [1]

The groups contained 7,668, 7,668, and 7,667 participants, respectively. [1]

Cumulative weekly drops in listening engagement. [1]

Regular messages alone, and regular messages with staff calls allocated in rotation. [1]

Engagement drops fell by 32.0% versus regular messages alone and 28.3% versus round-robin calling. [1]

The comparison with round-robin calling favored the model, but its p-value was .098. The comparison with messaging alone had p = .044. [1]

The study measured listening, not clinical health. Cumulative drops can count the same participant in more than one week. [1]

Each calling group received about 225 staff calls a week for seven weeks. [1]

Public SAHELI code and a data dictionary cover the later deployed system. They do not identify the software release used in the 2021 trial. [1] [5]

The study reports a limit on staff calls, not a monetary benefit. [1]

No cash amount is reported for the benefit or delivery cost. [1]

No monetary estimate is reported. [1]

Resources were allocated weekly over a seven-week study. [1]

The engagement improvement was observed. No monetary saving was measured. [1]

ARMMAN supplied the staff calls. The paper does not report their cost. [1]

No gross or net financial benefit was calculated. [1]

The resource constraint was staff capacity: about 225 calls per week in each calling group. [1]

The study compared calling policies at similar call capacity, rather than comparing cash expenditure. [1]

The authors did not assign a monetary value to the engagement improvement. [1]

No return or payback estimate is reported. [1]

ARMMAN ran mMitra and supplied health workers. The research partners developed and evaluated the scheduler with the NGO. [1]

The field study took place in ARMMAN’s mMitra program in India. [1]

ARMMAN operated the health program and held its participant records. Research partners used anonymized data. [1]

The program needed telephone access, listening logs, registration records, and staff who could place live calls. [1]

Researchers were based in India, the United States, and Singapore. The paper does not identify where participant data were hosted or transferred. [1]

Not assessed

The scheduler needs usable listening histories, current registration details, and a known limit on staff calls. [1]

A 2022 deployment study also found fewer engagement drops. A 2024 study reported some better health behaviors among selected survey respondents. [4] [5]

SAHELI entered routine use in April 2022. Later studies evaluated revised models and health-behavior outcomes. [4] [5]

The developers reported continuous SAHELI use from April 2022. [5]

Not assessed

Not assessed

Not assessed

Not assessed

AIBD-0099-source-0001, AIBD-0099-source-0002, AIBD-0099-source-0003, AIBD-0099-source-0004, AIBD-0099-source-0005 [1] [2] [3] [4] [5]

The published seven-week trial supplies the engagement result. Dated developer reports cover deployment. The 2025 preprint is the original source for later health-behavior findings. [1] [2] [3] [4]

English. [1] [2] [3] [4]

The sources come from the development and implementation partnership. Repeated trial results are not independent corroboration. [1] [2] [3] [4]

The 2022 paper reports the trial. The 2023 deployment paper links code. The 2026 chapter’s health-behavior findings originate in the 2025 preprint. [1] [2] [3] [4] [5]

The original studies, deployment account, SAHELI chapter sections, and linked code documentation were read within the scopes recorded in the source notes. [1] [2] [3] [4] [5]

Selecting live-call recipients with the model helped sustain listening within the available staff capacity. [1]

The sources are linked to the findings that they support. [1] [2] [3] [4] [5]

A peer-reviewed randomized study supports the engagement finding. The authors developed the system and included the implementing NGO. [1]

Not assessed

Continued listening and effective allocation of staff calls. [1]

Personal: 2; Organizational: 2 [1]

The study found a measurable improvement in sustained access to health information, without establishing clinical health gains. ARMMAN improved engagement within its limited capacity for staff calls. [1]

Observed in a randomized field study. [1]

Not assessed

Not assessed

Not assessed

The makers' claim [1] [2] [3] [4] [5]

A randomized field study of call scheduling. [1]

Optum's algorithm choosing which patients get extra-care outreach under-predicted Black patients' needs; ARMMAN's bandit likewise selects which mothers receive limited service calls.

Optum Algorithmic Health Risk Scores Reportedly Underestimated Black Patients' Needs

Maternal and child health information. [1]

The system selected mothers for limited staff calls to encourage continued listening to mMitra messages. [1]

The seven-week study began on 26 April 2021. [1]

ARMMAN tested the scheduler in its operating mMitra program. [1]

SHA-256: 95fb8c572187…

AI Incident Database: 124.

Not assessed

Not assessed

8 revisions. Latest: 25 October 2026.