Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation
Authors: Konstantin Hess, Valentyn Melnychuk, Dennis Frauen, Stefan Feuerriegel
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 NUMERICAL EXPERIMENTS Data: We benchmark our method using the established pharmacokinetic-pharmacodynamic tumor growth model by Geng et al. (2017). ... Performance metrics: We use different metrics to assess whether the uncertainty estimates of our method are faithful and sharp. ... 5.2 RESULTS Faithfulness: We first evaluate whether the estimated Cr Is are faithful. ... Sharpness: We compare the median width of estimated Cr Is (see Fig. 3). ... Error in point estimates: We further compare the errors in the Monte Carlo mean estimates of the treatment effects. ... 5.3 COMPARISON OF MODEL UNCERTAINTY |
| Researcher Affiliation | Academia | Konstantin Hess, Valentyn Melnychuk, Dennis Frauen & Stefan Feuerriegel Munich Center for Machine Learning LMU Munich {k.hess,melnychuk,frauen,feuerriegel}@lmu.de |
| Pseudocode | No | The paper describes its method using mathematical equations and architectural diagrams (Fig. 1), but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | 1https://github.com/konstantinhess/Bayesian-Neural-CDE |
| Open Datasets | Yes | Data: We benchmark our method using the established pharmacokinetic-pharmacodynamic tumor growth model by Geng et al. (2017). ... The tumor data was simulated according to the lung cancer model by Geng et al. (2017), which has previously been used in (Lim et al., 2018; Bica et al., 2020; Li et al., 2021; Melnychuk et al., 2022; Seedat et al., 2022; Vanderschueren et al., 2023). |
| Dataset Splits | Yes | For training, validation, and testing, our observed time window is 55 days with an additional prediction window of up to 5 days. For training and validation, we simulated 10, 000 and 1000 observations, respectively. For testing, we simulated the trajectories and the outcomes for 10, 000 patients under both treatment arms, respectively. |
| Hardware Specification | Yes | Experiments were carried out on 1 NVIDIA A100-PCIE-40GB. |
| Software Dependencies | No | The paper mentions 'Optimizer Adam (Kingma & Ba, 2015)' but does not specify version numbers for other key software components or libraries like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | Table 3: Hyperparameters of BNCDE and TE-CDE. ... Batch size 64 ... Max. number of epochs 500 ... MC samples training 10 ... Hidden state size dz 8 ... Learning rate 10^-3 ... Hidden layers 2 ... Hidden dimensions (128, 128) ... Diffusion coefficient 0.001 |