This blog is the high-level summary of the talk given by Prof. Milind Tambe, Harvard/Google AI at the PathCheck DICE Global Health Innovators seminar! Check out https://dice.pathcheck.org/talks.html to know more about previous talks.
The talk was focused on various efforts on AI for Social Impact — Wildlife Conservation, Healthcare, Public Safty, and more. In particular, Prof. Tambe discussed his work with Google Research India on using Restless Multi-Armed Bandits (RMABs) to provide better follow-up for expecting (and new) mothers.
In addition, Prof. Tambe also discussed various other efforts on — using influence maximization in graph networks for bringing awareness of STDs in homeless youth of Los Angeles (we had Prof. Bryan Wilder present this work in our earlier episode — https://youtu.be/QwWCpZkhU04); using constrained optimization methods to understand how to effectively deploy rangers to protect wildlife against poachers and much more! All of these are tied together under the broad umbrella of Multiagent Systems research.
The current state of Maternal and Childcare in India — a woman dies every 15 mins; 4/10 children are too thin/short. Prof. Tambe and his team partnered with ARMMAN, a non-profit in India, to provide 2 min reminder calls to expecting and new mothers about child/maternal care. Here, the non-profit has limited caller agents and a far higher number of mothers to contact. Therefore, it is important that each call done by the agent is effective.
The team made use of the whittle index to decide which arm to pull in RMABs (to select which mother to contact) to maximize the throughput. The Whittle index is calculated based on a fixed set of features available about the mothers (including their age, occupation, etc) and these features are further passed through a representation learning model to calculate the whittle index value. The authors noticed that the usage of RMABs brought in statistically significant results than the round-robin and no-contact baselines.
Dr. Aparna Hegde, Founder of ARMMAN on the collaboration — “The AI pilot with Google Research India helps us identify women at the highest risk of dropping out of the mMitra program and implement targeted interventions to improve engagement. When women listen to the mMitra calls on a regular basis, it leads to better health outcomes for both mothers and children. AI is helping us save lives by bringing women back into the fold and I am excited as ARMMAN scales the model to more women enrolled in the mMitra program.”
Please check out the complete video and the paper if you would like to know more details!
Title:Field Study in Deploying Restless Multi-Armed Bandits: Assisting Non-Profits in Improving Maternal and Child Health
Authors: Aditya Mate, Lovish Madaan, Aparna Taneja, Neha Madhiwalla, Shresth Verma, Gargi Singh, Aparna Hegde, Pradeep Varakantham, Milind Tambe
Abstract: The widespread availability of cell phones has enabled non-profits to deliver critical health information to their beneficiaries in a timely manner. This paper describes our work to assist non-profits that employ automated messaging programs to deliver timely preventive care information to beneficiaries (new and expecting mothers) during pregnancy and after delivery. Unfortunately, a key challenge in such information delivery programs is that a significant fraction of beneficiaries drop out of the program. Yet, non-profits often have limited health-worker resources (time) to place crucial service calls for live interaction with beneficiaries to prevent such engagement drops. To assist non-profits in optimizing this limited resource, we developed a Restless Multi-Armed Bandits (RMABs) system. One key technical contribution in this system is a novel clustering method of offline historical data to infer unknown RMAB parameters. Our second major contribution is evaluation of our RMAB system in collaboration with an NGO, via a real-world service quality improvement study. The study compared strategies for optimizing service calls to 23003 participants over a period of 7 weeks to reduce engagement drops. We show that the RMAB group provides statistically significant improvement over other comparison groups, reducing ~ 30% engagement drops. To the best of our knowledge, this is the first study demonstrating the utility of RMABs in real world public health settings. We are transitioning our RMAB system to the NGO for real-world use.
Bonus — Prof. Tambe offers his roadmap for those working on AI for Social Impact summarized in the image below!
Q&A Highlights
How do you handle the data-drift that might happen after deploying the model?
Currently, the models are occasionally re-trained. The cadence of retraining is decided by a set of experts based on evaluating the model performance variations. The team is interested to try other methods that enable online learning or continuously developing the model.
Can a similar constrained optimization model be used for vaccine allotment or to help other aspects in the current (or future) pandemics?
An important distinction would be that the state space is not observable in the case of a pandemic, thereby making it more difficult. That said, Prof. Tambe’s team had previously worked on collapsing bandit problem to tackle the case of providing care in a constrained environment where the state of the patients is not known.
What is the most important aspect when deploying health applications and how do we collect data?
The important aspect is to work with local communities. In our project for homeless people in LA we worked with various volunteer organizations and for the childcare/maternal care program, we partnered with ARMAAN which has in total 2.2 million women in their mMitra program (the calling-based reminder program), and overall they service 25 million women. This enables communities to have trust in the system. In general, working with local communities will help us understand the problem better. Further, Prof. Tambe added that it is of utmost importance that we as CS researchers visit the field and understand for whom we are building the solutions.
I would like to thank both our speaker — Prof. Milind Tambe — and my amazing co-organizers — Ramesh Raskar, Nina Reščič, Shanice Hudson, and Tavpritesh Sethi!
Also, if you’re interested in joining our efforts at PathCheck Foundation, head over to — tiny.cc/pathcheckslack. Feel free to drop any feedback or let me know if you would like to attend/speak at our event!
You can reach out to me at https://rsk97.github.io