Risk-Averse Active Sensing for Timely Outcome Prediction under Cost Pressure

Authors: Yuchao Qin, Mihaela van der Schaar, Changhee Lee

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our method outperforms baseline active sensing approaches in experiments with both synthetic and realworld datasets, and we illustrate the significance of our policy decomposition and the necessity of a risk-averse sensing policy through case studies. In the experiments, we evaluate the effectiveness of our proposed risk-averse active sensing approach RAS on both synthetic and real-world healthcare datasets.
Researcher Affiliation Academia Yuchao Qin Univerisity of Cambridge, UK Mihaela van der Schaar University of Cambridge, UK Alan Turing Institute, UK Changhee Lee Chung-Ang University, South Korea
Pseudocode Yes Algorithm 1 Risk-averse active sensing
Open Source Code Yes The source code of RAS can be found in the two Git Hub repositories listed below: The van der Schaar lab repo: https://github.com/vanderschaarlab/cvar_sensing The author s personal repo: https://github.com/yvchao/cvar_sensing
Open Datasets Yes ADNI dataset. The Alzheimer s Disease Neuroimaging Initiative2 (ADNI) dataset includes records on AD progression of N = 1002 patients with regular follow-ups every six months.2https://adni.loni.usc.edu
Dataset Splits Yes We first fit the outcome estimator f P on each dataset with 64/16/20 train/validation/test splits
Hardware Specification No The paper does not specify the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions using 'neural CDE' and 'Python package torchcde for interpolation' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes In our experiments, the hyperparameters of each active sensing method are selected based on the cost efficiency, i.e., ROC / Cost, obtained on the test set. For the synthetic data considered in our manuscript, we set the minimum and maximum allowed acquisition intervals as min = 0.2, max = 1.0, respectively. ... RAS: the coefficient for diagnostic error λ = 300 {200, 250, 280, 300, 310, 320, 350, 400}, discount factor γ = 0.99, tail-risk quantile α = 0.1, penalty for invalid visits ν = 10. All methods are trained with K = 200 iterations in the experiments. For RAS, we set the tail subset update interval M = 10 for Algorithm 1 in the manuscript.