Source-linked AI summary
Deep Learning for Patient-Specific Kidney Graft Survival Analysis
Margaux Luck, Tristan Sylvain, Héloïse Cardinal, Andrea Lodi, Yoshua Bengio
TL;DR
Estimating an individual patient’s survival time is challenging under conventional survival-model assumptions. The paper proposes a deep learning approach that predicts survival over time while jointly learning event-time and ranking objectives. It reports higher C-index performance than prior state-of-the-art methods, including a traditional Cox comparison on the study dataset.
Problem
Conventional models can have difficulty precisely estimating individual survival times, and the Cox proportional hazards assumption restricts feature effects to remain constant over time.
Source: §1 · pp. 1–2Method
The model directly predicts survival probabilities over time while jointly using event-time prediction and Cox partial log-likelihood ranking objectives.
Source: §1 · p. 2, §3 · p. 3, §3.1 · p. 3Results
C-index was 0.655 for the proposed method versus 0.65 for the traditional Cox model on this dataset, and the authors report outperforming previous state-of-the-art methods on C-index.
Source: §4.4 · p. 5, §5 · p. 7Takeaways & Limitations
The authors conclude that directly modeling the survival function offers advantages and provides clues about survival predictions at different time thresholds.
Source: §5 · p. 7Cox-model survival predictions are not directly possible without additional assumptions because its baseline hazard is not directly estimated.
Source: §2.2 · p. 3
Abstract
from arXiv · showhide
An accurate model of patient-specific kidney graft survival distributions can help to improve shared-decision making in the treatment and care of patients. In this paper, we propose a deep learning method that directly models the survival function instead of estimating the hazard function to predict survival times for graft patients based on the principle of multi-task learning. By learning to jointly predict the time of the event, and its rank in the cox partial log likelihood framework, our deep learning approach outperforms, in terms of survival time prediction quality and concordance index, other common methods for survival analysis, including the Cox Proportional Hazards model and a network trained on the cox partial log-likelihood.
1 Introduction
Survival data are often censored, and conventional models have limitations when estimating an individual patient’s survival time. The paper introduces a multi-task deep learning method and reports better C-index results than previous state-of-the-art methods.
1 Introduction
- Censoring occurs when an event is unobserved during the study, so the last visit provides only a lower bound on survival time.Source: §1 · p. 1
- Conventional survival models can struggle to estimate an individual patient’s survival time because their chosen distributions may limit prediction accuracy.The Cox proportional hazards assumption also constrains feature effects to remain constant over time, which the authors describe as restrictive in practice.Source: §1 · pp. 1–2
- Prior work includes stratified Cox models, multi-task logistic regression, and machine-learning approaches such as neural networks, Bayes classifiers, and decision trees.The passage cites neural-network comparisons with linear Cox models and a broader survival-analysis framework.Source: §1 · p. 2
- The method jointly predicts event time and rank in the Cox partial log-likelihood framework, and reports better C-index results than previous state-of-the-art methods.The authors evaluate it on real-life survival datasets.Source: §1 · p. 2
2 Background: Survival time analysis
This section defines right-censored survival data, describes established survival models and their constraints, and explains the concordance index used to evaluate ranking performance.
2.1 Notations
- Right-censored data include cases where event time is unobserved; tied event times complicate comparisons and require modifying the loss function.The survival function gives the probability of surviving beyond t, while the hazard is the instantaneous event probability conditional on no earlier event.Source: §2.1 · p. 2
2.2 Linear models: Cox proportional hazards model
- Cox models use a baseline hazard and feature effects, but survival predictions are not directly possible without extra assumptions because the baseline hazard is not directly estimated.The proportional hazards model is described as useful for characterizing existing cases rather than commonly predicting new ones.Source: §2.2 · pp. 2–3
2.3 Non-linear models
- Nonlinear alternatives include Random Survival Forests, Deep survival, and deep exponential families for modeling event time.Deep survival parametrizes the hazard with a neural network [9]; the paper’s Cox-loss implementation differs in its batch optimization, tie handling, and hazard parameterization.Source: §2.3 · p. 3
2.4 Standard evaluation: the concordance index
- The C-index measures how often a model correctly ranks admissible survival-time pairs, accommodates right censoring, and ranges from 0.5 for random prediction to 1 for perfect prediction.A pair is inadmissible when both events are censored or the earlier time is censored; tied pairs count as half correct.Source: §2.4 · p. 3
3 A deep learning survival model
The model directly predicts graft-failure time and survival probabilities using a deep network. It combines ranking and time-prediction losses, with evaluation adapted to censored data and tied events.
3 A deep learning survival model
- The model directly computes graft-failure time and predicts the probability of being alive at each study time, accounting for censored events.This is framed as an ordinal-regression-like task adapted for censoring.Source: §3 · p. 3
3.1 Model
- A patient-feature network uses a scalar score for Cox partial log-likelihood ranking and T sigmoid outputs for survival probabilities across time.The second loss penalizes incorrectly predicted times; censored output units receive no loss or back-propagated error, as illustrated in Figure 1.Source: §3.1 · pp. 3–4, Figure 1 · p. 4
3.2 Handling ties and censored data
- The cost combines Cox partial likelihood with Efron’s approximation for ties and a censored-data-adjusted ranking loss based on acceptable pairs.The time-prediction loss uses a convex loss, specified here as squared distance.Source: §3.2 · p. 4
3.3 Evaluation procedure
- Evaluation uses C-index and AUROC as pairwise-ranking measures, modifies AUROC for right censoring, and selects hyperparameters and stopping using a validation subset.The data split preserves uncensored-patient proportions and event proportions per time-step in training, test, and validation sets.Source: §3.3 · pp. 4–5
4 Experiments
The experiments use deceased-donor kidney transplants from the SRTR, preprocess clinical and biological features, and tune model hyperparameters by random search.
4.1 Dataset
- The dataset contains 131,709 deceased-donor transplants from 2000–2014; recipients under 18 and simultaneous multi-organ transplants were excluded.Graft failure means return to dialysis, retransplant, or death; 86,104 patients (around 65%) were censored.Source: §4.1 · p. 5
4.2 Preprocessing
- Features with more than 20% missing values were removed, and remaining missing values were filled using medians for continuous features and modes for categorical features.One-hot encoding and unit scaling produced 436 features between 0 and 1.Source: §4.1 · p. 5, §4.2 · p. 5
4.3 Hyper-parameters and training
- Hyperparameters were selected by random search, with batch size strongly affecting the partial log-likelihood loss because it uses pairwise combinations within each batch.Training used Adam, dropout, batch normalization, L1 and L2 regularization, and gradient clipping.Source: §4.3 · p. 5