Multimodal Prototyping for cancer survival prediction

Authors: Andrew H. Song, Richard J. Chen, Guillaume Jaume, Anurag Jayant Vaidya, Alexander Baras, Faisal Mahmood

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive evaluation on six cancer types shows that our framework outperforms state-of-the-art methods with much less computation while unlocking new interpretability analyses.
Researcher Affiliation Academia 1Mass General Brigham, Boston, MA, USA. 2Harvard Medical School, Boston, MA, USA. 3Massachusetts Institute of Technology, Cambridge, MA, USA. 4Johns Hopkins University School of Medicine, Baltimore, MD, USA.
Pseudocode No The paper provides detailed explanations and mathematical derivations but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes The code is available at https: //github.com/mahmoodlab/MMP.
Open Datasets Yes We use publicly available The Cancer Genome Atlas (TCGA) to evaluate MMP across six cancer types: Bladder urothelial carcinoma (BLCA) (n = 359), Breast invasive carcinoma (BRCA) (n = 868), Lung adenocarcinoma (LUAD) (n = 412), Stomach adenocarcinoma (STAD) (n = 318), Colon and Rectum adenocarcinoma (CRC) (n = 296), and Kidney renal clear cell carcinoma (KIRC) (n = 340). ... Log-2 transformed transcripts per million bulk RNA sequencing expression for all TCGA cohorts is accessed through UCSC Xena database (Goldman et al., 2020).
Dataset Splits Yes Following standard practice, we use 5-fold site-stratified cross-validation to mitigate batch effect (Howard et al., 2021).
Hardware Specification No The paper does not explicitly state the specific hardware (e.g., GPU models, CPU types, or memory) used to run the experiments.
Software Dependencies No The paper mentions optimizers (Adam W) and libraries in citations (e.g., scikit-survival), but does not provide a reproducible description of ancillary software with specific version numbers for key components like programming languages or deep learning frameworks.
Experiment Setup Yes All models are trained with a 1 10 4 learning rate with cosine decay scheduler, Adam W optimizer, and 1 10 5 weight decay for 20 epochs. MMP uses the Cox loss with a batch size of 64.