Simultaneous Prediction Intervals for Patient-Specific Survival Curves

Authors: Samuel Sokota, Ryan D'Orazio, Khurram Javed, Humza Haider, Russell Greiner

IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we consider four survival datasets. The Northern Alberta Cancer Dataset (NACD; 2402 patients, 53 features, 36% censorship), includes patients with many types of cancer: including lung, colorectal, head and neck, esophagus, stomach, etc. The other three datasets are from The Cancer Genome Atlas (TCGA) Research Network [Broad Institute TCGA Genome Data Analysis Center, 2016]: Glioblastoma multiforme (GBM; 592 patients, 12 features, 18% censorship), Rectum adenocarcinoma (READ; 170 patients, 18 features, 84% censorship), and Breast invasive carcinoma (BRCA; 1095 patients, 61 features, 86% censorship). These datasets contain rightcensored patients patients for whom the dataset specifies only a lower-bound on survival time.
Researcher Affiliation Academia Samuel Sokota , Ryan D Orazio , Khurram Javed , Humza Haider and Russell Greiner Department of Computing Science, University of Alberta {sokota, rdorazio, kjaved, hshaider, rgreiner}@ualberta.ca
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/ssokota/spie.
Open Datasets Yes The other three datasets are from The Cancer Genome Atlas (TCGA) Research Network [Broad Institute TCGA Genome Data Analysis Center, 2016]: Glioblastoma multiforme (GBM; 592 patients, 12 features, 18% censorship), Rectum adenocarcinoma (READ; 170 patients, 18 features, 84% censorship), and Breast invasive carcinoma (BRCA; 1095 patients, 61 features, 86% censorship).
Dataset Splits Yes Each dataset was shuffled and divided into a training set and a testing set with a 75/25 split. For its greedy hill climbing procedure, GSPIE divided each test patient s survival curve samples into an optimization set and a validation set with a 50/50 split.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU, CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions using MTLR and NUTS, but it does not specify software dependencies with version numbers (e.g., specific Python library versions or software package versions) required for replication.
Experiment Setup Yes We trained MTLR using a regularization factor of 1/2, which was tuned with five-fold cross-validation and N time points (where N is the size of the training set), such that an equal number of events occurs between each two time points, as is recommended by [Yu et al., 2011]. We considered a Bayesian approach we collected 10,000 posterior samples from the posterior p(m | D) using NUTS [Hoffman and Gelman, 2014] and estimated joint prediction intervals for the predictive posterior distribution p(s(x, m) | D) of test set patients.