calico/impulse: Fit Impulse and Sigmoidal Curves to Longitudinal Data Using TensorFlow

Implements the phenomenological kinetic model of Chechik and Koller <doi:10.1089/cmb.2008.13TT> using Bayesian priors to improve interpretability. Two models can be fit: a sigmoidal model parameterized by a half-max time constant, an asymptote and a rate constant, as well as an impulse model which adds a second sigmoidal response described by a second time constant and asymptote. Priors enforce non-negativity of timing and rate coefficients and with appropriate tuning, focus support on plausible parameter ranges. TensorFlow is used to optimize the maximum posterior estimate (MAP) as a combination of a non-linear least squares likelihood and priors on kinetic coefficients.

install.packages("remotes")
remotes::install_github("calico/impulse")