Thesis Defense | Jingjing Tang| September 29, 2026 | 12:30pm
CBD and CPCB are proud to announce the following thesis defense:
Title: Reliable Epidemic Intelligence from Delayed and Revised Public Health Surveillance Data
Jingjing Tang
Tuesday September 29th @ 12:30pm
GHC 6501
Committee:
Roni Rosenfeld, Chair, CMU
Bryan Wilder, Co-chair, CMU
Harry Hochheiser, PITT
Cecile Viboud, NIH
Abstract:
Epidemic situational awareness depends on timely and accurate information about ongoing transmission. In practice, recent surveillance observations are incomplete because of reporting and administrative delays, while even finalized records remain delayed and noisy measurements of underlying epidemic activity. This thesis examines how these limitations shape epidemic inference, from forecasting incomplete surveillance records to characterizing the information they ultimately contain.
We first develop Delphi-RF, a nonparametric revision-forecasting framework that captures reporting and revision dynamics through designed features and generates distributional forecasts of mature surveillance values. The framework accommodates different data types, reporting frequencies, and diseases. We then extend this framework to AutoDelphiRF, which automates revision forecasting through pre-forecast diagnosis, adaptive pooling of location–lag tasks with similar revision behavior, and post-forecast calibration and reliability assessment, without requiring dataset-specific exploratory analysis.
Even after revisions have resolved, epidemiological delays constrain inference about latent epidemic dynamics. We characterize these inferential limits through a spectral analysis of observation-level identifiability, showing how delays attenuate temporal variation. We further characterize how this attenuation, together with observation noise, affects the distinguishability of short-timescale epidemic trajectories and epidemiologically relevant features such as short-term growth and the magnitude and duration of transient events.
Together, these contributions provide practical and efficient methods for real-time revision forecasting from incomplete surveillance data while characterizing the limits on what epidemic dynamics can be reliably inferred from delayed and noisy observations.
Roni Rosenfeld, Chair, CMU
Bryan Wilder, Co-chair, CMU
Harry Hochheiser, PITT
Cecile Viboud, NIH
Abstract:
Epidemic situational awareness depends on timely and accurate information about ongoing transmission. In practice, recent surveillance observations are incomplete because of reporting and administrative delays, while even finalized records remain delayed and noisy measurements of underlying epidemic activity. This thesis examines how these limitations shape epidemic inference, from forecasting incomplete surveillance records to characterizing the information they ultimately contain.
We first develop Delphi-RF, a nonparametric revision-forecasting framework that captures reporting and revision dynamics through designed features and generates distributional forecasts of mature surveillance values. The framework accommodates different data types, reporting frequencies, and diseases. We then extend this framework to AutoDelphiRF, which automates revision forecasting through pre-forecast diagnosis, adaptive pooling of location–lag tasks with similar revision behavior, and post-forecast calibration and reliability assessment, without requiring dataset-specific exploratory analysis.
Even after revisions have resolved, epidemiological delays constrain inference about latent epidemic dynamics. We characterize these inferential limits through a spectral analysis of observation-level identifiability, showing how delays attenuate temporal variation. We further characterize how this attenuation, together with observation noise, affects the distinguishability of short-timescale epidemic trajectories and epidemiologically relevant features such as short-term growth and the magnitude and duration of transient events.
Together, these contributions provide practical and efficient methods for real-time revision forecasting from incomplete surveillance data while characterizing the limits on what epidemic dynamics can be reliably inferred from delayed and noisy observations.