TLDR: A field trial of a collaborative bandit algorithm within India’s Kilkari mobile maternal health program significantly improved call pick-up rates by learning individual mothers’ preferred call times. This personalized scheduling approach, tested with approximately 6500 participants, demonstrated a statistically significant increase in message delivery compared to the traditional random calling method, showcasing the potential of machine learning to enhance large-scale public health interventions.
Mobile health (mHealth) programs have emerged as a powerful tool for disseminating crucial health information, particularly to underserved communities. In India, the Kilkari program stands as a testament to this, delivering vital maternal health information through weekly voice calls to over 10 million registered mothers. These messages cover essential topics like iron and calcium supplementation, antenatal care, and postnatal practices, aiming to improve maternal health outcomes across the country.
However, the effectiveness of such large-scale programs hinges on successful message delivery. Currently, Kilkari employs a random call scheduling strategy, often leading to missed calls and reduced message delivery, despite multiple re-attempts. This approach limits the reach and impact of critical health information, as it doesn’t account for individual mothers’ preferred call times.
To address this significant challenge, a recent field trial introduced a collaborative bandit algorithm designed to optimize call timing by learning individual mothers’ preferred call times. This innovative approach aims to harness similarities among mothers to jointly learn their preferences, making the learning process more efficient than trying to learn each mother’s preferences in isolation. The algorithm works by iteratively learning from user responses and interactions, adapting to individual preferences to maximize engagement.
The field trial was conducted in the Kalahandi and Puri districts of Odisha, India, involving approximately 6500 Kilkari participants. These participants were randomly assigned to either a Control Group, which continued to receive calls using the standard random calling strategy, or a Treatment Group, where calls were scheduled using the collaborative bandit algorithm. The study consisted of two phases: a three-week Baseline Phase where both groups received random calls to establish a baseline and collect data, followed by a two-week Intervention Phase where the collaborative bandit algorithm was deployed for the Treatment Group.
The results of the trial demonstrated a statistically significant improvement in call pick-up rates for the group using the collaborative bandit algorithm. Specifically, for beneficiaries in the ‘Mid Tier’ (those who don’t always pick up or never pick up calls), the treatment group achieved a call pick-up rate of 37.63%, compared to 35.55% in the control group. This represents a 5.83% improvement, highlighting the algorithm’s effectiveness in optimizing call scheduling. Analysis across different time slots also showed positive improvements, with one particular time slot (16:45:00 to 18:45:00) seeing an 11.61% increase in pick-up rate.
Also Read:
- Streamlining Ride-Sharing Dispatch with One-Step AI Optimization
- Machine Learning Enhances Operating Room Scheduling with Predictive Confidence
This research underscores the efficacy of personalized scheduling in mobile health interventions and highlights the immense potential of machine learning to improve maternal health outreach at scale. The findings have significant implications for public health policy and practice in India and beyond, suggesting that personalized call scheduling can significantly enhance the reach and impact of maternal health programs. The study was a joint effort between ARMMAN, a non-profit in India, and Google Deepmind India, with beneficiary data fully anonymized to ensure privacy and security. For more details, you can refer to the full research paper here.
