Improving Health Information Access in the World's Largest Maternal Mobile Health Program via Bandit Algorithms

Arshika Lalan, Shresth Verma, Paula Rodriguez Diaz, Panayiotis Danassis, Amrita Mahale, Kumar Madhu Sudan, Aparna Hegde, Milind Tambe, Aparna Taneja

Introduction

Mobile health (mHealth) programs have emerged as a critical tool for delivering life-saving information to underserved populations. India’s Kilkari program stands as the world’s largest maternal mHealth initiative, delivering Automated Voice Messages (AVMs) to over 3 million active subscribers across 18 states. Despite its scale, the program faces a significant "leaky bucket" problem: listenership often declines over time, leading to beneficiaries being dropped from the system to conserve bandwidth. To address this, researchers developed CHAHAK, an AI-driven system designed to optimize engagement through non-Markovian Time-Series Bandits.

The Challenge: Scale, Scarcity, and Memory

Previous state-of-the-art solutions for mHealth engagement, such as the SAHELI system, relied on Markovian Restless Multi-Armed Bandits (RMAB). However, Kilkari presents unique challenges that render traditional Markovian models insufficient:

  1. Massive Scale: Managing interventions for millions of users requires extreme computational efficiency.
  2. Data Scarcity: Unlike smaller pilot programs, Kilkari lacks rich demographic profiles, relying almost exclusively on call logs.
  3. Non-Markovian Dynamics: Human behavior in long-term programs is rarely "memoryless." A user’s likelihood of engaging today is heavily influenced by their long-term history rather than just their most recent state.
  4. Resource Constraints: High-impact interventions, such as home visits by Accredited Social Health Activists (ASHA), are extremely limited and must be allocated with surgical precision.

The CHAHAK Framework

CHAHAK (meaning "chirp" or "glow" in Hindi) optimizes two distinct aspects of the mHealth pipeline: when to call and who to help.

1. Time-Slot Optimization

To maximize the probability of a call being answered, CHAHAK utilizes a Stochastic Bandit approach using the Upper Confidence Bound (UCB) algorithm. By modeling different times of day as "arms," the system learns the preferred time slots for individual beneficiaries—accounting for factors like work schedules or the availability of a shared family mobile phone.

2. Intervention Planning via MA-TARI

To prevent dropouts, the system must decide which beneficiaries should receive limited interventions (e.g., ASHA visits or automated reminders). CHAHAK moves beyond Markovian assumptions by employing:

  • LSTM-based Forecasting: A Long Short-Term Memory (LSTM) network predicts future listenership based on historical engagement patterns.
  • MA-TARI Index: The researchers introduced the Multi-Action Time-series Arm Ranking Index. This metric calculates the "lift" provided by an intervention—specifically the ratio of expected engagement duration with an intervention versus without one.
  • Dual Optimization: The system supports both Integer Linear Programming (ILP) for theoretical optimality and a Greedy approach for the computational efficiency required for national-scale deployment.

Data and Methodology

The system was developed and validated using real-world, anonymized data from the Odisha state in India. This included call logs from 4,000 beneficiaries registered in early 2022 and data from a June 2023 pilot campaign involving ASHA visits and automated reminders across six blocks.

One of the study's key insights was the "History Premium." The researchers demonstrated that a 6th-order Markov model (looking back six weeks) provided a 15% improvement in behavior prediction over a standard 1st-order model, confirming that long-term history is vital for modeling human engagement.

Results and Impact

In simulations and pilot analysis, CHAHAK demonstrated substantial improvements over baseline allocation methods:

  • Operational Efficiency: The UCB time-slot optimizer reduced the average number of call attempts required for a successful connection from 4.86 to 2.5 per week.
  • Increased Exposure: The MA-TARI policy led to a 57% increase in the total hours of health content consumed by beneficiaries compared to random allocation.
  • Retention Gains: The system successfully prevented 33% of predicted dropouts by the eighth week of simulation.
  • Scalability: The Greedy version of the MA-TARI algorithm achieved performance nearly identical to the ILP version while remaining computationally viable for millions of users.

Conclusion

CHAHAK represents a significant shift in mHealth optimization by moving from memoryless Markovian models to time-series-based bandits that respect the complexity of human behavior. By balancing sophisticated forecasting with practical, low-resource deployment strategies, the system offers a path toward drastically reducing dropouts in maternal health programs. Future work will focus on a national-scale deployment and a formal cohort study to further validate these gains in the field, potentially transforming how health information is delivered to millions of mothers across India.

Summary updated

BibTeX
@inproceedings{lalan2024improving,
  author = {Arshika Lalan and Shresth Verma and Paula Rodriguez Diaz and Panayiotis Danassis and Amrita Mahale and Kumar Madhu Sudan and Aparna Hegde and Milind Tambe and Aparna Taneja},
  title = {{Improving Health Information Access in the World's Largest Maternal Mobile Health Program via Bandit Algorithms}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30329},
}
Figure