An AI scheduling system deployed across a maternal health programme in India reduced drops in beneficiary engagement by 31% over five years, with measurable improvements in real-world health behaviours among new mothers, according to a study published on ArXiv in April 2025.

The SAHELI project (2020–2025), a collaboration between academic researchers and Indian NGO ARMMAN, set out to solve a practical and urgent problem: too few healthcare workers, too many vulnerable mothers, and no reliable way to decide who most needs a check-in call. The project applied a class of AI decision-making algorithms known as Restless Multi-Armed Bandits (RMABs) to allocate those scarce human resources as effectively as possible.

Why Standard Scheduling Fails Vulnerable Mothers

In large-scale public health programmes, attrition is the enemy. When beneficiaries stop engaging — missing calls, skipping check-ins — the preventive care that could save lives goes undelivered. The traditional approach to this problem, known as "predict-then-optimise" or Two-Stage modelling, first builds a model to predict who might disengage, then uses a separate system to decide who gets a call. The two steps are developed independently, which means errors in prediction compound into poor allocation decisions.

SAHELI's key methodological shift was abandoning that two-step process in favour of Decision-Focused Learning (DFL) — an approach that trains the AI model with the actual goal of maximising engagement built directly into the learning process, rather than treating prediction and decision-making as separate problems.

The DFL policy reduced cumulative engagement drops by 31% relative to the current standard of care, significantly outperforming the Two-Stage model.

The distinction matters in a resource-constrained setting. A system optimised purely for predictive accuracy might correctly identify at-risk mothers but still allocate calls inefficiently. DFL, by contrast, learns what a good scheduling decision looks like — not just what a good prediction looks like.

What the Randomised Controlled Trial Found

The research team tested the system through large-scale randomised controlled trials, the gold standard for evaluating real-world interventions. Mothers enrolled in ARMMAN's mMitra programme — which delivers preventive health information via mobile phone — were assigned to groups receiving AI-optimised outreach or standard care.

The SAHELI-DFL system outperformed both the standard of care and the older Two-Stage RMAB model on engagement metrics. Critically, the study did not stop at engagement data. Researchers also measured whether the increased contact translated into actual health behaviour changes — and found it did. Mothers in the AI-optimised group showed statistically significant improvements in continued consumption of iron and calcium supplements, nutrients essential to maternal and infant health, particularly in the period following birth.

Iron deficiency anaemia affects an estimated 40% of pregnant women in India according to the National Family Health Survey, making supplement adherence a direct indicator of health risk reduction, not merely a proxy metric.

Five Years from Research to Scalable Implementation

The SAHELI project is notable not only for its results but for its duration. Most AI health interventions are evaluated in short pilots. A five-year deployment running from 2020 through 2025 allowed the research team to observe how the system performs as beneficiary populations change, how models degrade or adapt over time, and how a real NGO with real capacity constraints integrates AI into daily operations.

ARMAN's mMitra programme serves hundreds of thousands of beneficiaries across India. Operating at that scale means the system's decisions have direct human consequences — a misconfigured algorithm does not just produce a worse test score, it means a high-risk mother does not receive a call she needed.

The RMAB framework is mathematically well-suited to this kind of problem. It models each beneficiary as an independent "arm" with a hidden internal state — engaged or disengaged — that transitions probabilistically based on whether they receive an intervention. The system must decide, at each time step, which arms to "pull" given a fixed budget of healthcare worker calls. The Restless variant of the problem allows states to evolve even when no action is taken, which reflects the real dynamic: mothers can disengage between calls without any system interaction.

Limits and What Comes Next

The paper, published as a preprint on ArXiv and not yet peer-reviewed in a journal, describes the project as providing a "scalable blueprint" for applying sequential decision-making AI to health resource allocation. The authors acknowledge that different health contexts have different state dynamics, different intervention types, and different data availability constraints.

The DFL approach also requires sufficient historical engagement data to train effectively, which may limit transferability to programmes with shorter track records or smaller beneficiary pools. Whether the 31% reduction in engagement drops holds across different geographies, languages, or health programme types remains an open empirical question that the authors do not resolve.

Nevertheless, the methodology is theoretically transferable. Programmes targeting tuberculosis adherence, childhood vaccination schedules, or chronic disease management face structurally similar resource allocation problems — limited outreach workers, heterogeneous beneficiary risk, and the need to make sequential decisions under uncertainty.

What This Means

For health programme designers and AI researchers alike, SAHELI demonstrates that closing the loop between prediction and decision-making — not just building a smarter model, but building one aligned with what the model is actually for — can translate directly into lives improved, not just benchmarks beaten.