Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning Survival Distributions with the Asymmetric Laplace Distribution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a parametric survival model that predicts per-individual Asymmetric Laplace Distribution parameters (location, scale, asymmetry) by maximum likelihood with censoring, giving closed-form summaries.
Problem
Survival models either impose restrictive parametric distributions or use nonparametric discretized/quantile outputs (DeepHit, CQRNN) that are computationally heavy, piecewise, and sensitive to censoring pseudo-value and grid choices.
Approach
Model each individual's event time as an Asymmetric Laplace Distribution whose location, scale, and asymmetry are predicted by a neural network, trained by censored maximum likelihood (density for observed events, survival function for censored), avoiding CQRNN's pseudo-value re-weighting and quantile grid.
Key results
- Across 189 comparisons (21 datasets, 9 metrics), the proposed ALD model is significantly better than CQRNN in 23% of comparisons and worse in 12%; vs LogNorm 60% better / 5% worse; vs DeepHit 73% better / 4% worse.
- The ALD model achieves better IBS than all baselines on nearly every dataset (e.g., 21 of 21 vs LogNorm, DeepSurv, DeepHit).
- Performance is weaker on the SUPPORT dataset, where calibration is comparable to baselines but not superior, attributed to the ALD assumption and skewed event distribution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the ALD survival model instead of CQRNN | Survival prediction · 21 datasets, IBS | IBS: 19 better, 1 worse, 1 same | Helped |
| Use the ALD survival model instead of LogNorm | Survival prediction · 21 datasets, IBS | IBS: 21 better, 0 worse, 0 same | Helped |
| Use the ALD survival model instead of DeepSurv | Survival prediction · 21 datasets, IBS | IBS: 21 better, 0 worse, 0 same | Helped |
| Use the ALD survival model instead of DeepHit | Survival prediction · 21 datasets, IBS | IBS: 21 better, 0 worse, 0 same | Helped |
| Use the ALD survival model instead of CQRNN | Survival prediction · 21 datasets, MAE | MAE: 6 better, 8 worse, 7 same | Mixed |
| Use the ALD survival model instead of LogNorm | Survival prediction · 21 datasets, MAE | MAE: 10 better, 3 worse, 8 same | Mixed |
| Use the ALD survival model instead of DeepSurv | Survival prediction · 21 datasets, MAE | MAE: 6 better, 8 worse, 7 same | Mixed |
| Use the ALD survival model instead of DeepHit | Survival prediction · 21 datasets, MAE | MAE: 12 better, 6 worse, 3 same | Mixed |
| Use the ALD survival model instead of CQRNN | Survival prediction · 21 datasets, Harrell's C-Index | Harrell's C-Index: 4 better, 2 worse, 15 same | Mixed |
| Use the ALD survival model instead of DeepHit | Survival prediction · 21 datasets, Harrell's C-Index | Harrell's C-Index: 15 better, 0 worse, 6 same | Helped |
| Use the ALD survival model instead of CQRNN | Survival prediction · 21 datasets, Uno's C-Index | Uno's C-Index: 2 better, 3 worse, 16 same | Mixed |
| Use the ALD survival model instead of DeepHit | Survival prediction · 21 datasets, Uno's C-Index | Uno's C-Index: 15 better, 0 worse, 6 same | Helped |
Surprising findings
- Using the mean as summary does not beat CQRNN on C-index and MAE, and the paper notes these gains are not consistent across summaries.
- The approach is trained on a MacBook CPU, indicating light compute requirements.
Limitations
- Relies on the Asymmetric Laplace Distribution assumption, which may not fit all datasets (e.g., SUPPORT calibration).
- Choosing the appropriate summary statistic (mean, median, mode, quantile) for a dataset is non-trivial.
- Mean summary does not outperform CQRNN on C-index and MAE.
Experimental setup
Datasets: METABRIC WHAS SUPPORT GBSG TMBImmuno BreastMSK LGGGBM 14 synthetic datasets with synthetic censoring
Benchmarks: METABRIC WHAS SUPPORT GBSG TMBImmuno BreastMSK LGGGBM 14 synthetic datasets
Metrics: MAE Integrated Brier Score Harrell's C-index Uno's C-index CensDcal calibration slope and intercept for S(t|x) and f(t|x)
Baselines: LogNorm MLE DeepSurv DeepHit CQRNN
Builds on: CQRNN DeepHit DeepSurv LogNorm MLE
Closest papers by meaning
- Survival Analysis via Density Estimation (ICML 2025 · similarity 0.86)
- Deep Copula-Based Survival Analysis for Dependent Censoring with Identifiability Guarantees (AAAI 2024 · similarity 0.85)
- Toward Conditional Distribution Calibration in Survival Prediction (NeurIPS 2024 · similarity 0.84)
- Learning Survival Models with Right-Censored Reporting Delays (NeurIPS 2026 · similarity 0.84)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.83)
- Assessing the robustness of heterogeneous treatment effects in survival analysis under informative censoring (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.