Continuous-Time Modeling of Counterfactual Outcomes Using Neural Controlled Differential Equations

Authors: Nabeel Seedat, Fergus Imrie, Alexis Bellot, Zhaozhi Qian, Mihaela van der Schaar

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we validate the ability of TE-CDE to estimate counterfactual outcomes from irregularly sampled observational data. [...] TE-CDE consistently outperforms existing approaches in all simulated scenarios with irregular sampling.
Researcher Affiliation Academia 1Department of Applied Mathematics and Theoretical Physics, University of Cambridge, UK 2University of California, Los Angeles, USA 3Columbia University, USA 4The Alan Turing Institute, London, UK.
Pseudocode No The paper describes the methods textually and mathematically but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes https://github.com/seedatnabeel/TE-CDE and https://github.com/vanderschaarlab/mlforhealthlabpub/tree/main/alg/TE-CDE
Open Datasets Yes First, we describe a simulation environment based on a Pharmacokinetic Pharmacodynamic (PK-PD) model of lung cancer tumor growth (Geng et al., 2017), which allows counterfactuals to be calculated at any time point for arbitrary treatment plans. [...] We consider four treatment options: no treatment, chemotherapy, radiotherapy, and combined chemotherapy and radiotherapy. The assignment of chemotherapy and radiotherapy are modeled as Bernoulli random variables with probabilities pc and pr, respectively, that depend on tumor diameter3 as follows: pc(t) = σ γc Dmax D(t) θc , pr(t) = σ γr Dmax D(t) θr , where Dmax = 13cm is the maximum tumor diameter, θc = θr = Dmax/2 and D(t) is the average tumor diameter.
Dataset Splits Yes Unless otherwise stated, each experiment is run with 10,000 patients for training, 1,000 for validation and 10,000 for testing.
Hardware Specification Yes The model was implemented with Pytorch and Torch CDE and was trained and evaluated on a single Nvidia P100 or T4 GPU.
Software Dependencies No The paper mentions PyTorch, torchcde, torchdiffeq, and the Dormand Prince (dopri5) solver, but does not provide specific version numbers for these software components.
Experiment Setup Yes The integrand of the neural CDE (i.e. f) is a 2-layer neural network with hidden states of size=128. The dimensionality of the latent state z is 8. We use linear interpolation when defining the control path Xt. [...] For domain adversarial training, we use the standard procedure (Ganin et al., 2016), with an initial µ = 0 that follows an exponentially increasing schedule per epoch of training for the range [0, 1]. [...] Both encoder and decoder are trained for 100 epochs each. That said we also include early stopping in the training protocol based on the validation loss, with patience=5. When MC Dropout is included, we use a dropout probability=0.1. [...] We did tune the learning rate (lr = 1e-3, 1e-4 , 1e-5, 1e-6) based on performance on the validation set.