Multimodal Patient Representation Learning with Missing Modalities and Labels

Authors: Zhenbang Wu, Anant Dadu, Nicholas Tustison, Brian Avants, Mike Nalls, Jimeng Sun, Faraz Faghri

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate MUSE on three publicly available datasets: MIMIC-IV, e ICU, and ADNI. Results show that MUSE outperforms all baselines, and MUSE+ further elevates the absolute improvement to 4% by extending the training scope to patients with absent labels.
Researcher Affiliation Collaboration Zhenbang Wu1,2,3 , Anant Dadu2,3, Nicholas Tustison4, Brian Avants5, Mike Nalls2,3, Jimeng Sun1, Faraz Faghri2,3 1University of Illinois Urbana-Champaign, 2National Institutes of Health 3Data Tecnica International, 4University of Virginia, 5University of Pennsylvania
Pseudocode Yes The pseudocode of MUSE is available in Appx. A. Algorithm 1: Training and Inference for MUSE.
Open Source Code Yes The code of MUSE is publicly available 1. 1https://github.com/zzachw/MUSE
Open Datasets Yes We evaluate MUSE on three publicly available datasets: MIMIC-IV (Johnson et al., 2023), e ICU (Pollard et al., 2018), and ADNI (Jack et al., 2008).
Dataset Splits Yes We split the dataset into 70%, 10%, 20% training, validation, and test sets.
Hardware Specification Yes We report the per-epoch training time and AUC-ROC score for the MIMIC-IV mortality prediction task on a single NVIDIA A100 GPU. 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 MUSE using Py Torch Paszke et al. (2019) 1.11 and Python 3.8.
Experiment Setup Yes The edge dropout rate is set to 15%. We train all models for 100 epochs on the training set, and select the best model by monitoring the performance on the validation set. The final results are reported on the test set. The cosine similarity temperature τ is set to 0.05. Specifically, we only tune the learning rate for both MUSE and the baseline methods while keeping the other hyperparameters aligned to ensure a fair comparison.