NeuralCohort: Cohort-aware Neural Representation Learning for Healthcare Analytics
Authors: Changshuo Liu, Lingze Zeng, Kaiping Zheng, Shaofeng Cai, Beng Chin Ooi, James Wei Luen Yip
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness and generalizability of Neural Cohort are validated across extensive real-world EHR datasets. Experimental results demonstrate that Neural Cohort consistently improves the performance of various backbone models, achieving up to an 8.1% increase in AUROC. |
| Researcher Affiliation | Academia | 1National University of Singapore, Singapore 2Zhejiang University, China 3National University Hospital, Singapore. Correspondence to: Changshuo Liu <EMAIL>. |
| Pseudocode | No | The paper describes the methodology in prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain an explicit statement or a direct link indicating that the source code for the proposed Neural Cohort method is publicly available. |
| Open Datasets | Yes | We conduct extensive experiments across three widely-recognized real-world EHR datasets: MIMIC-III (Johnson et al., 2016), MIMIC-IV (Johnson et al., 2020) and Diabetes130 (Clore John & Beata, 2014) |
| Dataset Splits | Yes | We randomly split each dataset into training, validation, and test sets with a ratio of 8: 1: 1. |
| Hardware Specification | Yes | All the experiments are conducted on a server with Intel(R) Xeon(R) W-2133 CPU @ 3.60GHz, 64G memory, and 3 NVIDIA Ge Force RTX 2080 Ti. |
| Software Dependencies | Yes | In the Pre-context Cohort Synthesis Module, we employ icdcodex (Fisher, 2020) and Glo Ve (Pennington et al., 2014) as gp and gs in the Path and Semantics components of Hierarchical Visit Engine to distinguish similar paths in the ontology tree and capture the semantics, respectively. |
| Experiment Setup | Yes | The dimensions of embedding for diagnosis, medication and laboratory test, visit and patient is 128, 64, 64, 128 and 128. The dimensions of R0, RL, RG, and Rfinal are the same dimension as the setting of the backbone. λpcs, λJS and λco are set to 0.1 in our experiments. We use the Adam optimizer with an initial learning rate of lr = 1e 3. Hyperparameters Kp = 5, |C| = 1100 remain consistent across all datasets and downstream tasks. |