Speaker
Description
Although ordinary and partial differential equation (ODE/PDE) models are widely used in infectious disease inference and public health decision-making, parameter sensitivity and identifiability are often assessed with tools that lack mechanistic interpretability and overlook sensitivity to initial conditions or derived epidemiological quantities. We present a framework for generating interpretable measures of sensitivity and identifiability by computing forward sensitivities, elasticities, and singular value decomposition (SVD) diagnostics of sensitivity structure. Using a susceptible-exposed-infected-recovered model as an example, we augment the system with sensitivity equations using the Jacobian matrix, incorporate an observation mapping from the full (unknown) state space to observed data, and treat initial conditions as parameters. We evaluate sensitivities of both model states and derived latent quantities. Sensitivities and elasticities identify distinct periods of parameter importance, while SVD diagnostics reveal shifting dominant parameter combinations and when inference is stronger or weaker during an outbreak under different observation models. We extend the framework to a PDE vector-borne disease model and a serocatalytic model, demonstrating scalability and the impact of observation design on identifiability. This workflow supports model calibration, experimental design, and early diagnosis of practical identifiability issues.
