The study also shows why donor heterogeneity matters. Donations from individuals in economically disadvantaged communities declined even though those communities were among the hardest hit by the crisis and prioritized immediate, time-sensitive needs. At the same time, small and local institutions played an important role in supporting high-need schools. These findings suggest that crowdfunding platforms can improve crisis-time resource mobilization by distinguishing between individual and institutional donors and using more targeted engagement strategies. Read the paper.
]]>We develop a diagnostic system for healthcare that allocates mammography interpretation tasks between AI algorithms and radiologists. Rather than viewing the choice as a simple comparison between full automation and human-only decision making, the framework considers a delegation strategy in which AI evaluates cases, makes recommendations for low- and high-risk cases, and sends ambiguous cases to radiologists for additional assessment.
A key result is an analytically derived two-threshold policy. When AI performance does not exceed that of radiologists, the system can recommend no follow-up for low-risk cases, recommend follow-up for high-risk cases, and delegate cases in an intermediate risk range to human experts. The optimal thresholds depend on the economics and safety tradeoffs of the diagnostic system rather than on the system’s prior history.
Backtesting with data from a mammography AI contest and real-world cost and performance measures shows potential cost savings of up to 20.9% compared with an expert-alone approach. More broadly, the paper argues that the important organizational question is not whether machines will simply replace people, but how decision-making tasks should be divided between humans and machines to achieve better outcomes at sustainable cost.
The article is an advance online publication in Production and Operations Management.
]]>Using screening mammography as the application, we develop an optimization framework that compares three approaches: an expert-alone strategy in which radiologists interpret mammograms, an automation strategy in which AI performs the task, and a delegation strategy in which AI evaluates cases and selectively refers appropriate cases to radiologists. The analysis shows that the best approach depends not only on predictive performance, but also on disease prevalence, the relative costs of false-positive and false-negative decisions, algorithm costs, expert performance, and potential liability. In backtesting with data from the Digital Mammography DREAM Challenge, the delegation strategy was optimal for the strongest-performing algorithm and produced substantial cost savings compared with relying solely on human experts. More broadly, the results suggest that the future of AI-enabled healthcare may often involve carefully designed human–AI workflows rather than a simple choice between humans and full automation.
The article was published in Nature Communications, volume 16, article 2289, on March 7, 2025.
]]>The study uses a discrete-event simulation of a high-volume breast imaging center and incorporates AI-based triage together with risk information from the Tyrer-Cuzick model and a deep-learning risk model. We compare alternative scheduling templates and examine how they affect daily recalls, workflow congestion, and the use of same-day diagnostic capacity. The results show how scheduling decisions can be designed around patient risk rather than treating every screening appointment in the same way.
More broadly, the work highlights an important connection between predictive analytics and operations management in healthcare: improving outcomes does not depend only on developing better risk models, but also on designing workflows that use those predictions effectively. Risk-stratified scheduling can help align diagnostic resources with patient needs and reduce unnecessary follow-up visits.
The paper was published in Journal of the American College of Radiology, volume 22, issue 3, pages 297–306, in March 2025.
]]>However, the challenge lies in efficiently quantifying the effects of these combinations, given the vast number of possible combinations and the resources required. Our recent research introduces a novel framework—Correlated Drug Action (CDA)—to address this challenge.
Combination therapies can be studied at two levels: in vitro on cells and in vivo on living organisms. In vitro research focuses on dose response at a fixed time post-drug administration (dose-space models), while in vivo research focuses on survival time at fixed doses (temporal models).
Traditionally, null models have established a baseline for expected drug combination effects. These models are essential for determining if a combination is more effective than expected.
Building on the principle of Independent Drug Action (IDA), which posits that each drug in a combination acts as if the other drug were absent, we introduce the temporal Correlated Drug Action (tCDA) model. Unlike previous models with time-varying correlation coefficients, tCDA employs a non-time-varying coefficient, offering a fast and scalable solution.
The tCDA model describes the effect of a combination based on individual monotherapies and a population-specific correlation coefficient. This model is valid for generic joint distributions of survival times characterized by their Spearman correlation.
We applied the tCDA model to public oncology clinical trial data involving 18 different combinations. The model effectively explained the effect of clinical combination therapies and identified combinations that could not be explained by tCDA alone. When the survival distribution of a combination is explained by tCDA, the estimated correlation parameter can reveal sub-populations that may benefit more from one monotherapy or the combination.
To address the limitations of translating preclinical cell line results to clinical outcomes, we adapted IDA’s temporal-space ideas to dose-space, resulting in the dose-space CDA (dCDA) model. This model describes the effect of combinations in cell cultures in terms of the dosages required for each monotherapy to kill cells after treatment.
The dCDA model estimates the correlation between joint distribution dosages, similar to the tCDA and ORR models in patient cohorts. Using MCF7 breast cancer cell line experiments, we demonstrated dCDA’s effectiveness in assessing potential drug synergy. We introduced the Excess Over CDA (EOCDA) metric to evaluate possible synergy, allowing for non-zero correlations.
]]>Our paper published in PNAS proposes an innovative approach: mapping the probability of correct classification to the probability of fermion occupation in a quantum system, specifically using the Fermi–Dirac distribution. This novel perspective facilitates calibrated probabilistic outputs and introduces new methodologies for optimizing classification thresholds and evaluating classifier performance.
At its core, the Fermi–Dirac distribution describes the statistical distribution of particles over energy states in systems obeying Fermi-Dirac statistics, typically applied to fermions that adhere to the Pauli exclusion principle. In this paper, we adapt the mathematical form of this distribution to model the probability of correct classification in binary classifiers.
By leveraging this quantum analogy, we establish a framework where:
We analyzed the effects of policy changes that reduce the number of Airbnb listings using a difference-in-difference approach. Our findings indicate that such self-regulations lead to a reduction in overall crime rates. Notably, incidents of assault, robbery, and burglary decreased, although theft incidents saw an increase.
To understand how these effects vary across different neighborhoods, we employed geographically weighted regression. Our analysis revealed that socioeconomic factors like income, housing prices, and population density significantly moderate the impact of Airbnb occupancy on crime rates. This highlights the importance of local context in evaluating the outcomes of platform self-regulations.
Our research provides empirical evidence on the societal impacts of the sharing economy and the role of platform self-regulation. By showing how reducing home-sharing listings can influence crime rates, and how these effects differ across neighborhoods, our findings offer valuable insights for policymakers. These insights can help shape regulations that promote both the benefits of the sharing economy and community safety.
]]>In our recent paper, we study and try understand in depth how the molecular response of cells to each of two drugs combine when the two drugs are given in combination. Check it out here.
]]>With the availability of new algorithms almost every day and the heterogeneity of the data, it might be a wise approach to base our predictions not on a single algorithm but rather on an ensemble of algorithms. SUMMA achieves this in an unsupervised way. More details can be found here.
]]>In this work, we developed a novel algorithm, NeTFactor, that uses a computationally-inferred context-specific gene regulatory network and applies topological, statistical, and optimization methods to identify a minimal set of regulators underlying a biomarker set. The paper can be found here.
]]>