Hamsa Bastani

Angel Tsai-Hsuan Chung, Botong Zhang, Ling-Chieh Kung, Hamsa Bastani, Osbert Bastani, Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning.

Abstract: Generative AI (GenAI) is rapidly reshaping education by unlocking the potential for personalized tutoring. Yet, emerging platforms largely focus on GenAI chatbot tutors that reactively answer student questions. We hypothesize that the efficacy of GenAI chatbot tutors can be substantially improved by proactively guiding student learning. To test this, we design a novel tutoring platform that tightly integrates a carefully-designed GenAI chatbot with a reinforcement learning algorithm for sequencing practice problems. Critically, this algorithm leverages rich signals from student-chatbot interactions to adaptively select practice problems of an appropriate difficulty level. In partnership with the Taipei City Government and American Institute in Taiwan, we deployed our tutoring platform in conjunction with a five-month course to teach Python to students across ten high schools. We randomized students between a fixed practice problem sequence and our adaptive sequencing algorithm. We find that adaptive sequencing increased unassisted final exam performance by 0.15 standard deviations (equivalent to 6-9 months of schooling by some estimates); mediation analysis suggests that gains were driven by increased engagement. Our work provides large-scale field evidence that student-chatbot interactions provide valuable signals for proactively optimizing and personalizing student learning.

Angel Tsai-Hsuan Chung, Jatu Abdulai, Patrick Bayoh, Lawrence Sandi, Francis Smart, Hamsa Bastani, Osbert Bastani (2026), Improving Access to Essential Medicines via Decision-Aware Machine Learning, Nature (Research Article) . 10.1038/s41586-026-10433-7

Abstract: A critical challenge in healthcare systems in low- and middle-income countries (LMICs) is the efficient and equitable allocation of scarce resources, particularly essential medicines. This problem is complicated by limited high-quality data, which restricts the applicability of traditional data-driven techniques. We propose a novel decision-aware machine learning framework for essential medicines allocation, which additionally leverages multi-task learning to ensure sample efficiency and catalytic priors to ensure equitable allocation. In collaboration with the Sierra Leone national government, we performed a staggered, nationwide deployment of our system as a decision support tool. Our econometric evaluation finds an estimated 19% increase in consumption of allocated products in treated districts, demonstrating its efficacy at improving access to essential medicines. Our tool was subsequently scaled nationwide, covering an estimated 2 million women and children under five. Our work demonstrates how machine learning methods can improve efficiency at very low cost in resource-constrained global health settings.

Haosen Ge, Neha Sharma, Hamsa Bastani, Osbert Bastani (Under Revision), Rethinking Algorithmic Fairness for Human-AI Collaboration.

Abstract: Existing approaches to algorithmic fairness aim to ensure equitable outcomes \emph{if} human decision-makers comply perfectly with algorithmic decisions. However, perfect compliance with the algorithm is rarely a reality or even a desirable outcome in human-AI collaboration. Yet, recent studies have shown that selective compliance with fair algorithms can \textit{amplify} discrimination relative to the prior human policy. As a consequence, ensuring equitable outcomes requires fundamentally different algorithmic design principles that ensure robustness to the decision-maker's (a priori unknown) compliance pattern. We define the notion of \textit{compliance-robustly} fair algorithmic recommendations that are guaranteed to (weakly) improve fairness in decisions, regardless of the human's compliance pattern. We propose a simple optimization strategy to identify the best performance-improving compliance-robustly fair policy.  However, we show that it may be infeasible to design algorithmic recommendations that are simultaneously fair in isolation, compliance-robustly fair, and more accurate than the human policy; thus, if our goal is to improve the equity and accuracy of human-AI collaboration, it may not be desirable to enforce traditional algorithmic fairness constraints. We illustrate the value of our approach on criminal sentencing data before and after the introduction of an algorithmic risk assessment tool in Virginia.

Hamsa Bastani, Osbert Bastani, Angel Tsai-Hsuan Chung, “Optimizing Health Supply Chains in LMICs with Machine Learning: A Case Study in Sierra Leone”. In Responsible and Sustainable Operations: The New Frontier, edited by Tang, Christopher S. (Switzerland: Springer Nature, 2024), pp. 187-202

Abstract: This chapter overviews the challenges in pharmaceutical supply chains (PSCs) in Low- and Middle-Income Countries (LMICs), with a focus on Sierra Leone. Furthermore, it describes how traditional supply chain optimization strategies can be used to improve performance of PSCs in Sierra Leone. Finally, it describes the significant potential for using machine learning in this framework for effective demand forecasting. We highlight challenges such as limited data availability, the need to ensure equitable distribution, as well as the potential for transfer learning to address some of these challenges.

Xinmeng Huang, Kan Xu, Donghwan Lee, Seyed Hamed Hassani, Hamsa Bastani, Edgar Dobriban Optimal Heterogeneous Collaborative Linear Regression and Contextual Bandits.

Angel Tsai-Hsuan Chung, Jimmy Qin, Ping-Cheng Lin, Hamsa Bastani, The Impact of Human-AI Collaboration in Outpatient Care: Evidence from Somaliland.

Abstract: Healthcare documentation is one of the major operational burdens in resource-constrained health systems, often competing directly with patient care. This paper evaluates a human-AI collaborative documentation tool implemented at the largest public hospital in Somaliland. Physicians enter a clinical note draft, from which AI generates a polished clinical note that physicians may directly submit, edit then submit, or abandon AI note and submit their own draft. Exploiting the staggered rollout of this tool across hospital departments, we estimate intent-to-treat (ITT) effects using a Callaway and Sant’Anna (2021) difference-in-differences design. Our results show that AI significantly reduces service time by 34% and reduces service time dispersions across encounters. These findings suggest that AI has the potential to increase effective service capacity and standardize care delivery. We find no evidence that these gains come at the expense of clinical care quality. In particular, provider surveys indicate greater patient interaction and clinical note quality improves substantially. However, AI-enabled faster service do not necessarily improve patient access and downstream operational benefits, for example, patient waiting time does not decrease significantly. We show that this is because physicians strategically respond to the reduced service time by shortening total working time and by starting clinical day later. Furthermore, we find significant improvement on physicians’ independently written notes (without AI) post-treatment, suggesting that improvement does not necessarily require adoption; exposure to AI (but without adoption) can be of great value as well.

Arielle Anderer, Hamsa Bastani, John Silberholz (2022), Adaptive Clinical Trial Designs with Surrogates: When Should We Bother?, Management Science (2022).

Abstract: The success of a new drug is assessed within a clinical trial using a primary endpoint, which is typically the true outcome of interest, e.g., overall survival. However, regulators sometimes approve drugs using a surrogate outcome --- an intermediate indicator that is faster or easier to measure than the true outcome of interest, e.g., progression-free survival --- as the primary endpoint when there is demonstrable medical need. While using a surrogate outcome (instead of the true outcome) as the primary endpoint can substantially speed up clinical trials and lower costs, it can also result in poor drug approval decisions since the surrogate is not a perfect predictor of the true outcome. In this paper, we propose combining data from both surrogate and true outcomes to improve decision-making within a late-phase clinical trial. In contrast to broadly used clinical trial designs that rely on a single primary endpoint, we propose a Bayesian adaptive clinical trial design that simultaneously leverages both observed outcomes to inform trial decisions. We perform comparative statics on the relative benefit of our approach, illustrating the types of diseases and surrogates for which our proposed design is particularly advantageous. Finally, we illustrate our proposed design on metastatic breast cancer. We use a large-scale clinical trial database to construct a Bayesian prior, and simulate our design on a subset of clinical trials. We estimate that our design would yield a 16% decrease in trial costs relative to existing clinical trial designs, while maintaining the same Type I/II error rates.

Kan Xu and Hamsa Bastani (Under Review), Learning Across Bandits in High Dimension via Robust Statistics.

Abstract: Decision-makers often face the "many bandits" problem, where one must simultaneously learn across related but heterogeneous contextual bandit instances. For instance, a large retailer may wish to dynamically learn product demand across many stores to solve pricing or inventory problems, making it desirable to learn jointly for stores serving similar customers; alternatively, a hospital network may wish to dynamically learn patient risk across many providers to allocate personalized interventions, making it desirable to learn jointly for hospitals serving similar patient populations. We study the setting where the unknown parameter in each bandit instance can be decomposed into a global parameter plus a sparse instance-specific term. Then, we propose a novel two-stage estimator that exploits this structure in a sample-efficient way by using a combination of robust statistics (to learn across similar instances) and LASSO regression (to debias the results). We embed this estimator within a bandit algorithm, and prove that it improves asymptotic regret bounds in the context dimension d; this improvement is exponential for data-poor instances. We further demonstrate how our results depend on the underlying network structure of bandit instances.

Hamsa Bastani, Pavithra Harsha, Georgia Perakis, Divya Singhvi (2021), Learning Personalized Product Recommendations with Customer Disengagement, MSOM.

Abstract: Problem definition: We study personalized product recommendations on platforms when customers have unknown preferences. Importantly, customers may disengage when offered poor recommendations. Academic/practical relevance: Online platforms often personalize product recommendations using bandit algorithms, which balance an exploration-exploitation trade-off. However, customer disengagement—a salient feature of platforms in practice—introduces a novel challenge because exploration may cause customers to abandon the platform. We propose a novel algorithm that constrains exploration to improve performance. Methodology: We present evidence of customer disengagement using data from a major airline’s ad campaign; this motivates our model of disengagement, where a customer may abandon the platform when offered irrelevant recommendations. We formulate the customer preference learning problem as a generalized linear bandit, with the notable difference that the customer’s horizon length is a function of past recommendations. Results: We prove that no algorithm can keep all customers engaged. Unfortunately, classical bandit algorithms provably overexplore, causing every customer to eventually disengage. Motivated by the structural properties of the optimal policy in a scalar instance of our problem, we propose modifying bandit learning strategies by constraining the action space up front using an integer program. We prove that this simple modification allows our algorithm to perform well by keeping a significant fraction of customers engaged. Managerial implications: Platforms should be careful to avoid overexploration when learning customer preferences if customers have a high propensity for disengagement. Numerical experiments on movie recommendations data demonstrate that our algorithm can significantly improve customer engagement.

Hamsa Bastani, Kimon Drakopoulos, Vishal Gupta (Forthcoming), Interpretable OR for High-Stakes Decisions: Designing the Greek COVID-19 Testing System.

Abstract: In the summer of 2020, in collaboration with the Greek government, we designed and deployed Eva – the first national scale, reinforcement learning system for targeted COVID-19 testing. In this paper, we detail the rationale for three major design/algorithmic elements: Eva’s testing supply chain, estimating COVID-19 prevalence, and test allocation. Specifically, we describe the design of Eva’s supply chain to collect and process thousands of biological samples per day with special emphasis on capacity procurement. Then, we propose a novel, empirical Bayes estimation strategy to estimate COVID-19 prevalence among different passenger types with limited data and showcase how these estimates were instrumental for a variety of downstream decision-making. Finally, we propose a novel, multi-armed bandit algorithm that dynamically allocates tests to arriving passengers in a non-stationary environment with delayed feedback and batched decisions. All of our design and algorithmic choices emphasize the need for transparent reasoning to enable human-in-the- loop analytics. Such transparency was crucial to building trust and buy-in among policymakers and public health experts in a period of global crisis.