Data-driven operations for health & human services

Working Paper. Huifeng Su, Lesley Meng, Edieal J. Pinker

Over 100,000 victims rely on self-reported, semi-structured clues on humanitarian service platforms to search for their missing parent or child through text comparisons. However, such searches face significant challenges, including vast search spaces, inaccurate data reporting, and complex matching patterns among text pairs. Despite its societal importance, the operational challenges inherent in these large-scale search efforts remain understudied. We analyze structured and unstructured data from a large online family-reunification forum and develop a novel deep learning–based recommendation system that effectively handles these challenges. Our approach significantly narrows search spaces and improves human search quality and efficiency, outperforming existing state-of-the-art solutions, including LLM-based approaches. We show that, when tailored and combined appropriately, smaller but domain-adapted models can deliver superior performance while remaining fast, cost-free, and locally deployable for non-profit organizations. Beyond directly enhancing search quality and efficiency, we demonstrate that text-comparison–based recommendations can also improve DNA collection compliance, creating new opportunities for the design and operation of family-reunification services.

Working Paper. Huifeng Su, Lesley Meng, Rohit Sangal, Edieal J. Pinker

We examine how disruptive critical incidents (CIs), defined as acute and high-intensity patient arrivals requiring resuscitation, affect emergency department (ED) physicians’ productivity and decision quality. Using four years of electronic medical record and administrative data from a large U.S. hospital and a quasi-experimental design, we find that, unlike prior studies, CI exposure does not reduce overall productivity. Physicians sustain throughput by multitasking, extending work hours, and increasing diagnostic testing to support disposition decisions. However, post-CI discharges exhibit higher 30-day revisit rates for conditions where diagnostic imaging typically provides limited clinical value. These findings underscore the need for system-level interventions to support physicians after CIs, particularly when decisions demand high cognitive effort.

Working Paper. Huifeng Su, Lesley Meng, Rohit Sangal, Edieal J. Pinker

Emergency Department (ED) boarding occurs when admitted patients wait in the ED for an inpatient bed to become available due to limited staffed beds, high demand, or hospital flow inefficiencies. Using an instrumental variable design and data from a large academic medical center, we estimate the causal impact of boarding on patient outcomes. Each additional hour of boarding increases hospital length of stay by 0.8%, the odds of care-level escalation by 16.7%, and total charges by 1.3%. The effect varies significantly by condition, acuity, and age, reflecting differences in the value-added care that the ED can provide during boarding. These findings on heterogeneous waiting costs highlight that patient prioritization can improve overall hospital throughput when inpatient resources are limited.

JAMA Network Open. (2023). Sangal RB, Su H, Khidir H, Parwani V, Liebhardt B, Pinker EJ, Meng L, Venkatesh AK, Ulrich A.  

In a cross-sectional study of 314,841 emergency department (ED) visits, 28.8% of patients experienced queue jumps, meaning they were overtaken by others who received care sooner, violating the acuity-based, first-come, first-served standard. Non-Hispanic Black, Hispanic or Latino, Spanish-speaking, and Medicaid-insured patients were more likely to be jumped over. Those who were jumped over had higher odds of being placed in hallway beds and of leaving before treatment was complete. These findings suggest that EDs should standardize triage and room assignment processes to mitigate bias and improve equity in patient access to timely emergency care.