Research map / Bandits and online learning / Contextual bandits and Thompson sampling
Improving Health Information Access in the World’s Largest Maternal Mobile Health Program via Bandit Algorithms
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CHAHAK, a bandit-based system for Kilkari (India's largest maternal mHealth program) that learns beneficiaries' preferred AVM call time slots with UCB and allocates ASHA visits and automated call reminders via a non-Markovian multi-action time-series bandit with an MA-TARI index.
Problem
Dwindling listenership and automated dropouts in Kilkari, a large-scale maternal mHealth voice-message program, under limited intervention resources and with no demographic features beyond listenership logs.
Approach
Replaces the Markovian RMAB used in prior SAHELI work with a non-Markovian time-series bandit that forecasts listenership with an LSTM and ranks arms per intervention via a recursive time-to-threshold index (MA-TARI), plus a per-beneficiary UCB for AVM time-slot selection.
Key results
- UCB reduces calls needed per week from 4.86 (uniform) to as few as 2.5 by week 6.
- MA-TARI LP and Greedy yield 49.6h and 49.47h additional listening by week 8 vs 31.5h for Random (about 57% more exposure).
- MA-TARI achieves 33% dropout prevention at the last week relative to control.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UCB instead of Uniform (no update) to choose AVM time slots | Kilkari AVM scheduling · Kilkari time-slot planning | calls per week until pickup: 4.86 -> 2.5 | Helped |
| use UCB rather than Uniform (update) to learn the best time slot | Kilkari AVM scheduling · Kilkari time-slot planning | convergence to best-slot pickup rate within difference ≤ 0.15: 4 weeks vs 8 weeks | Helped |
| allocate interventions with MA-TARI Integer Linear Programming rather than random selection | Kilkari intervention planning · Kilkari intervention simulation, 8 weeks | additional content listened (hours over control): 31.5h -> 49.6h | Helped |
| allocate interventions with MA-TARI Greedy rather than random selection | Kilkari intervention planning · Kilkari intervention simulation, 8 weeks | additional content listened (hours over control): 31.5h -> 49.47h | Helped |
| use MA-TARI Greedy instead of MA-TARI Integer Linear Programming (ablation) | Kilkari intervention planning · Kilkari intervention simulation | additional content listened (hours over control): 49.6h vs 49.47h | No clear effect |
| use LSTM rather than the other tested forecasting architectures (ablation) | Kilkari listenership forecasting · Kilkari forecasting test set | MAE: less than 0.01 | No clear effect |
| use ASHA visits rather than automated call reminders | Kilkari intervention campaigns · Kilkari intervention campaigns | listenership increase: 5% -> 16% | Helped |
Surprising findings
- Greedy MA-TARI performs on par with the ILP formulation.
- Relative to order-1 Markov, higher-order Markov models improve log-likelihood by up to 15%, supporting non-Markovian dynamics.
Limitations
- Evaluation is simulation-based using a fitted ground-truth simulator rather than a field trial.
- Demographic features unavailable, so only listenership history is used.
- Intervention budgets fixed at 1% per action type.
Experimental setup
Datasets: Kilkari anonymized call logs from Odisha (4000 beneficiaries) ARMMAN awareness campaign logs (June 2023, ASHA visits and automated calls)
Benchmarks: Kilkari Odisha call logs (simulation)
Metrics: average calls per week until pickup fraction of beneficiaries with pickup rate converged within cumulative additional listenership hours over control percentage of dropouts prevented mean absolute error of forecasting
Baselines: Uniform (no update) Uniform (update) Random Control group MA-TARI Greedy vs ILP
Builds on: restless multi-armed bandit SAHELI UCB
Closest papers by meaning
- Context in Public Health for Underserved Communities: A Bayesian Approach to Online Restless Bandits (AAAI 2025 · similarity 0.86)
- Optimizing Vital Sign Monitoring in Resource-Constrained Maternal Care: An RL-Based Restless Bandit Approach (AAAI 2025 · similarity 0.86)
- RoME: A Robust Mixed-Effects Bandit Algorithm for Optimizing Mobile Health Interventions (NeurIPS 2024 · similarity 0.84)
- A Decision-Language Model (DLM) for Dynamic Restless Multi-Armed Bandit Tasks in Public Health (NeurIPS 2024 · similarity 0.83)
- Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.81)
- PRIORITY2REWARD: Incorporating Healthworker Preferences for Resource Allocation Planning (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.