ManyDG: Many-domain Generalization for Healthcare Applications
Authors: Chaoqi Yang, M Brandon Westover, Jimeng Sun
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that Many DG can boost the generalization performance on multiple real-world healthcare tasks (e.g., 3.7% Jaccard improvements on MIMIC drug recommendation) and support realistic but challenging settings such as insufficient data and continuous learning. |
| Researcher Affiliation | Academia | 1University of Illinois Urbana-Champaign 2Harvard Medical School 3Beth Israel Deaconess Medical Center |
| Pseudocode | No | The paper describes the method using mathematical formulations and descriptive steps but does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/ycq091044/ManyDG. |
| Open Datasets | Yes | Seizure dataset is from Jing et al. (2018); Ge et al. (2021) for detecting ictal-interictal-injurycontinuum (IIIC) seizure patterns from electronic recordings. ... Sleep-EDF Cassette (Kemp et al., 2000) has in total 78 subject overnight recordings ... MIMIC-III (Johnson et al., 2016) is a benchmark EHR database ... The multi-center e ICU dataset (Pollard et al., 2018) contains more than 140,000 patients hospital visit records. |
| Dataset Splits | Yes | For each set of experiments in the following, we split the dataset into train / validation / test by ratio 80%:10%:10% under five random seeds and re-run the training for calculating the mean and standard deviation values. |
| Hardware Specification | Yes | The experiments are implemented by Python 3.9.5, Torch 1.10.2 on a Linux server with 512 GB memory, 64-core CPUs (2.90 GHz, 512 KB cache size each), and two A100 GPUs (40 GB memory each). |
| Software Dependencies | Yes | The experiments are implemented by Python 3.9.5, Torch 1.10.2 on a Linux server with 512 GB memory, 64-core CPUs (2.90 GHz, 512 KB cache size each), and two A100 GPUs (40 GB memory each). |
| Experiment Setup | Yes | More details, including backbone architectures, hyperparameters selection and datasets, are clarified in Appendix A.4. ... 128 as the batch size, 128 as the hidden representation size, 50 as the training epochs (50 size of an epoch size of an episode as the number of episodes for MLDG baseline, which follows MAML, the same setting for other datasets), 5e-4 as the learning rate with Adam optimizer and 1e-5 as the weight decay for all models. Our model uses τ = 0.5 as the temperature (the same for other datasets). |