An Iterative Self-Learning Framework for Medical Domain Generalization
Authors: Zhenbang Wu, Huaxiu Yao, David Liebovitz, Jimeng Sun
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the generalizability of SLDG across spatial and temporal data distribution shifts on two real-world public EHR datasets: e ICU and MIMIC-IV. Our results show that SLDG achieves up to 11% improvement in the AUPRC score over the best baseline. |
| Researcher Affiliation | Academia | Zhenbang Wu 1 Huaxiu Yao 2 David M Liebovitz 3 Jimeng Sun 1 1 University of Illinois Urbana-Champaign, {zw12, jimeng}@illinois.edu 2 University of North Carolina at Chapel Hill, huaxiu@cs.unc.edu 3 Northwestern University, david.liebovitz@nm.org |
| Pseudocode | Yes | The pseudocode of SLDG can be found in Appx. A. Algorithm 1: Training and Inference for SLDG. |
| Open Source Code | No | The paper does not contain any explicit statement about making their code publicly available, nor does it provide a link to a code repository for their implementation. |
| Open Datasets | Yes | We evaluate SLDG on two publicly available real-world EHR datasets: e ICU [38] and MIMIC-IV [18] |
| Dataset Splits | Yes | Each group is then split into 70% for training, 10% for validation, and 20% for testing. |
| Hardware Specification | Yes | The model is trained on a Cent OS Linux 7 machine with 128 AMD EPYC 7513 32-Core Processors, 512 GB memory, and eight NVIDIA RTX A6000 GPUs. |
| Software Dependencies | Yes | We implement SLDG using Py Torch [34] 1.11 and Python 3.8. For SLDG, UMAP [30] from UMAPlearn [41] is used with 2 components, 10 neighbors, and 0 minimum distance; and k-Means from Scikit-learn [35] is used with the default hyper-parameter. |
| Experiment Setup | Yes | All models are trained for 100 epochs, and the best model is selected based on the AUPRC score monitored on the source validation set. For SLDG, UMAP [30] from UMAPlearn [41] is used with 2 components, 10 neighbors, and 0 minimum distance; and k-Means from Scikit-learn [35] is used with the default hyper-parameter. We apply a dropout of rate 0.2. We use Adam as the optimizer with a learning rate of 1e-4 and a weight decay of 1e-5. The batch size is 256. |