Temporally-Consistent Survival Analysis

Authors: Lucas Maystre, Daniel Russo

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate empirically that it achieves better sample-efficiency and predictive performance compared to approaches that directly regress the observed survival outcome. In this section we evaluate TCSR (Algorithm 1) empirically on synthetic and real-world data.
Researcher Affiliation Collaboration Lucas Maystre Spotify lucasm@spotify.com Daniel Russo Columbia University & Spotify djr2174@gsb.columbia.edu
Pseudocode Yes Algorithm 1 Temporally-consistent survival regression (TCSR). Require: dataset of transitions T = {(xm, x m)}, initial parameters θ. 1: repeat 2: for m = 1, . . . , M do 3: ym1 1{x m = } Based on one-step observed outcome. 4: wm1 1 5: for k = 2, . . . , K do Based on predictions at x m. 6: ymk hθ(k 1|x m) 7: wmk Sθ(k 2|x m) 8: θ arg minθ PM m=1 PK k=1 wmk H[ymk hθ(k|xm)] 9: until θ has converged
Open Source Code Yes We provide a software library with a reference implementation of TCSR in the Python programming language as well as computational notebooks that enable reproducing the results presented in this section at https://github.com/spotify-research/tdsurv.
Open Datasets Yes Of the three datasets we consider, two contain real data from clinical studies and are publicly available online [30]. The first study tracks survival outcomes from 312 patients diagnosed with PBC, a rare liver disease [25]. The second study follows 467 patients diagnosed with HIV / AIDS [2].
Dataset Splits Yes We report the average performance obtained by using 5-fold cross-validation in Figure 2.
Hardware Specification Yes The experiments were conducted on a desktop computer with a 10th Gen Intel i9-10900X CPU (3.7 GHz) and two NVIDIA GeForce RTX 3090 GPUs.
Software Dependencies Yes In our implementation of TCSR, we delegate Line 8 in Algorithm 1 to a Newton-CG solver provided by the Sci Py library [31]. The referenced Sci Py version is 1.0.
Experiment Setup Yes For simplicity, we have fixed the number of iterations to 30 across all datasets and experiments, but the algorithm often converges in fewer iterations.