# Biyonka Liang Develops Model for Improving Resource Allocation Decisions in Public Health June 05, 2025 How can healthcare providers effectively decide which which patients to provide extra support when resources, such as budget and staff, are constrained? ARMMAN, an NGO in India that runs multiple large maternal and child health programs (interventions implemented by ARMMAN have reached over 60 million women now, including 3 million via mMitra and the rest through Kilkari) and their collaborators in Computer Science and Statistics at Harvard and at Google Research in India, have been driven to answer this question in their research to improve health outcomes for pregnant women. While all women enrolled in mMitra receive weekly automated calls with maternal healthcare information, ARMMAN wanted a mechanism to help allocate live calls to the women who would benefit the most from them. To address this problem, Biyonka Liang, Lily Xu, Milind Tambe, and Lucas Janson of Harvard University and Aparna Taneja of Google Research worked with ARMMAN to develop a model for allocating these live phone calls in a simulation study. The model, BCoR (Bayesian Learning for Contextual RMABs), introduces a novel approach by incorporating contextual information across beneficiaries to determine who would benefit the most from receiving a live call from a healthcare worker. To elucidate the significance of this new method and its potential applications, Biyonka Liang, a 2025 Statistics PhD graduate, met with us to discuss her work with her co-authors on their paper “Context in Public Health for Underserved Communities: A Bayesian Approach to Online Restless Bandits.” This work was recently published as part of the proceedings of AAAI, a major AI conference. ## Motivation of Public Health Collaboration Interested in research motivated by applied problems, particularly healthcare and genetics-related applications, Liang jumped on the opportunity to collaborate with colleagues in CS, Dr. Lily Xu (a PhD alum) and Professor Milind Tambe, who had an ongoing collaboration with Dr. Aparna Taneja and ARMMAN. More specifically, Liang, with Associate Professor Lucas Janson, was interested in bringing a statistical modeling approach to this resource allocation problem, both for ARMMAN and for other settings. Reflecting on her motivation, Liang said, “The goals of the project were to determine how to allocate resources effectively and to learn the type of person who is most responsive to an intervention. To answer these questions, you have to learn how people will behave from observed data, which is exactly what our statistical modeling contributed to this project.” Previous approaches to this problem often relied on more theoretical frameworks without considering the practical constraints of a real-world setting. These methods often lacked the ability to account for specific time parameters (such as the 40-week duration of pregnancy) and contextual factors (like education and income levels) that could influence a woman's likelihood of listening to and benefiting from the healthcare information provided in the calls. In addition, while there was an earlier real-world implementation for ARMMAN, the offline nature of the model prevented it from learning from data in real time. Reflecting on the importance of the collaboration with ARMMAN, Liang said, “It was very important to us throughout the project that we had ARMMAN’s input and contextual information so that the design of our model would directly address their concerns.” For example, for the simulation study, the nonprofit organization provided anonymized covariate data collected in 2022 from 24,011 enrolled women, including education level, income level, and phone ownership. This real-world data was crucial for incorporating context into the model and improving its ability to predict which women would benefit most from a live phone call. ## Unpacking the BCoR Method At a high level, BCoR (Bayesian Learning for Contextual RMABs) in the simulation study with ARMMAN works in the following way. During the first week, all women enrolled receive an automated call with information about how to sustain maternal and infant health. During the second week, the adherence to the treatment (whether women listened to the call) is observed and the algorithm uses this information and certain covariates to decide which women should receive a live call the next time. For example, if person A listened to the call, then she might be more likely to listen to calls in the future and may not need a live call. However, if person B did not listen to the phone call and is in a lower income and education bracket, then the algorithm uses this context to decide that the healthcare worker should call person B. In the third week, the algorithm tracks whether women A and B listened to the automated call and updates the probability of their future adherence. To describe the method at a more granular level, BCoR employs restless multi-armed bandits (RMABs). A traditional multi-armed bandit problem uses reinforcement learning algorithms to optimize rewards over time. Liang illustrated this concept: “Imagine that you are in a casino and there are multiple slot machines that will generate a different reward depending on your selection. Each slot machine represents an ‘arm’ and when you pull the arm, there is a probability that you will or won’t get money. Over time, from pulling different arms multiple times, you learn the reward distributions of the slot machines.” Restless multi-armed bandits, however, are more complex and often more applicable to real-life scenarios because the state of the arms, even for arms that aren’t pulled, and the reward distributions are in flux. In the simulation, the women beneficiaries are represented by the arms, pulling an arm means giving a woman a live call, and the reward is whether the woman listens to the automated call the next time. Liang explained the application of RMABs in their simulation: “Because Person A is already listening to automated calls, she will likely continue to do so, with or without a live call. On the other hand, if I give a live call to person B, who has not listened to automated calls, there is a chance that she will continue to not listen. The reward distributions for person A and B (probability of listening to the next automated call) depends on their current state, which means that the reward distribution isn’t fixed and can change over time.” ## Differentiating BCoR from Other Methods Previous theoretical work in RMAB models has assumed that you have the ability to observe people forever, without a time limit. However, in the ARMMAN example, women are only enrolled for about 80 weeks (throughout pregnancy and until one year after childbirth), which means that “you're often in a situation where your model has to determine how to allocate resources across the entire pool of beneficiaries, when you haven't even observed how the vast majority of them would respond to this type of intervention,” explained Liang. She further emphasized the advantage of BCoR: “The benefit of using BCoR is that you get much stronger empirical results when applying this method to finite time settings and small datasets from real life.” While other methods have attempted to address both resource allocation optimization and the identification of individuals most responsive to interventions, Liang highlighted a key difference: “Their approach is to treat every single arm as if it is independent from other arms, which assumes that there’s nothing learned from one arm that could be applied to another arm.” In contrast, BCoR leverages contextual data, such as demographic information, phone ownership, and adherence history, across all beneficiaries to inform its recommendations for live calls. This allows the model to even make informed recommendations for women new to the study. For example, if a woman in her third week of the study does not own a phone, and the model has substantial data indicating that women in a similar situation tend to decrease their adherence, BCoR would likely recommend prioritizing a live call for her. Summing up the value of BCoR, Liang stated, “Our work is different from existing work because our model learns continuously from data how people are going to behave with a level of greater granularity, while in the past, other methods assumed that everyone behaves the same across time.” ## Importance of Ethics in the Project Ethical considerations are paramount in public health research involving patient data. Liang and her co-authors explicitly addressed these concerns in their paper. They noted that the 24,011 women participants provided informed consent after receiving detailed information about the simulation study, including the anonymization process for their demographic data and ARMMAN’s data privacy policy. Liang and co-authors only had read-only access to the anonymized data for the purpose of this research project. Reflecting on the ethical considerations, Liang said, “While our model wasn’t yet deployed in real life, it was important to consider what the ethical implications would be for real-life deployment (since that is the ultimate goal).” She further clarified, “The purpose of our model was to determine who gets the stronger touch, based on who is at risk of not receiving the information; we never withheld or provided additional health information to women.” ## Future Exploration of Topics Liang expressed enthusiasm about the potential for this project to lead to new and exciting research directions. A key goal is to see BCoR implemented in real-world settings. Expanding on this idea, Liang said, “I’d love to see BCoR deployed in real life and work through the challenges, which would include challenges outside of statistics, such as logistics and computing resources.” There are also examples of other resource allocation problems, such as where to deploy park rangers to prevent anti-poaching, that Liang could potentially employ BCoR with collaborators and stakeholders. The work of Biyonka Liang and her collaborators demonstrates the powerful synergy between machine learning, statistical methods, and public health expertise in developing innovative solutions for critical resource allocation challenges. Their research holds the promise of improving health outcomes for underserved communities worldwide. ## Works Cited Liang, B., Xu, L., Taneja, A., Tambe, M., & Janson, L. (2025). “Context in Public Health for Underserved Communities: A Bayesian Approach to Online Restless Bandits.” Proceedings of the 39th AAAI Conference on Artificial Intelligence (AAAI). .