Yihan Hu

Purpose: Public-health forecasting is central to epidemic intelligence and operational decision support. In practice, surveillance data are affected by reporting delays, revisions, and backfill, as well as abrupt regime shifts, which often reduce model reliability across regions and systems. Methods: We developed EpiMap-LLM, a protocol-aware forecasting approach that links epidemic dynamics with surveillance context using a frozen language-model backbone and lightweight trainable components. Results: Across daily and weekly surveillance settings (JHU CSSE COVID-19 and CDC influenza hospitalization surveillance), EpiMap-LLM consistently improves MAE and RMSE over strong forecasting baselines. Conclusion: Protocol-aware forecasting improves robustness and practical usefulness for surveillance dashboards, early warning, and public-health decision support in heterogeneous reporting systems.

A conversational LLM clinical app that scores well on AUC is not the same as a clinical app whose explanations are grounded in real clinical guidance. We argue that LLM-based pediatric risk-assessment systems should be evaluated under two conditions — the current LLM-only setup and a retrieval-grounded setup using a fixed clinical corpus (CDC + WHO + IDSA guidance) — and scored on citation faithfulness, evidence-grounded correctness, and subgroup robustness alongside AUC. This is an architectural shift in how clinical LLM apps should be validated before deployment.

Recent work showed GPT-4 can map free-text breast pain descriptions to a binary triage recommendation with reasonable sensitivity. The end-to-end design has a structural problem: clinical reasoning for breast pain usually goes via intermediate attributes (focality, cyclicity, red-flag features) before risk stratification. An end-to-end LLM bypasses these auditable intermediate steps, which means failures cannot be localised. We argue clinical-LLM triage should be evaluated as a two-step pipeline (attribute extraction then risk stratification) with attribute-level metrics.

Content prevalence on algorithm-driven platforms is not a clean proxy for adolescent exposure. A TikTok feed is shaped by reach, engagement dynamics, and audience composition, not by aggregate posting patterns. We propose three cheap fixes for ongoing surveillance: view-weighted prevalence, bounded sensitivity analysis for missing nutrient data, and audits of multi-product appearances to estimate undercounting from one-product-per-video coding rules.

Most food-environment research and policy still assumes that exposure is shaped by neighbourhood retail — outlet density, school zones, buffers around homes. That model was right in 2010 but is incomplete in 2026: online food delivery platforms, social media food marketing, and algorithm-driven content curation are now major exposure pathways. We argue ultra-processed food policy needs an explicit digital track: platform-level marketing regulation parallel to broadcast rules, algorithmic exposure audits, and measurement frameworks that include view-weighted exposure rather than only retail proximity.

A theoretical companion to our Real Estate paper. We extend the Ngai-Tenreyro (2014) search-and-matching framework to monthly frequency, prove existence and uniqueness of the equilibrium, and calibrate to observed US data. Using SIPP household-mobility data and Google Trends as a corroborating indicator, we document a corresponding post-2021 shift in mobility timing. The calibrated model reproduces the spring shift in both prices and transaction volumes using mobility timing alone, without invoking changes in housing supply, credit conditions, or remote-work preferences.

Seasonality has traditionally shaped the US housing market, with activity peaking in spring-summer and declining in autumn-winter. This study uses X-13-ARIMA decomposition on FHFA Housing Price Index data and Census transaction data covering 1991 to 2024 and documents a clean structural break after 2020. Post-pandemic, the seasonal peak has moved to March/April; the amplitude of the cycle has also grown; the pattern is consistent across US regions.