🤖 AI

GRAFT: Decoupling Ranking and Calibration for Survival Analysis

The paper proposes GRAFT, a novel hybrid survival analysis model that decouples prognostic ranking from calibration by combining a linear AFT model with a non-linear residual neural network and stochastic feature selection, thereby achieving superior discrimination and calibration performance in high-dimensional, censored data settings.

Original authors: Mohammad Ashhad, Robert Hoehndorf, Ricardo Henao

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Ashhad, Robert Hoehndorf, Ricardo Henao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to predict how long a patient will live after a diagnosis. This is a tricky job because you often don't get to see the full story. Some patients leave the hospital early, or the study ends before they pass away. In statistics, this is called censoring. You know they survived at least until that point, but you don't know the exact end date.

For a long time, doctors have used two main tools to make these predictions:

The Old-School Calculator: These are simple, reliable formulas. They are great at telling you exactly how likely an event is to happen at a specific time (calibration), but they are rigid. They can only see straight lines and simple patterns. If the real world is messy and complex, these formulas get confused.
The Super-Computer: These are modern AI models. They are incredibly flexible and can spot complex, hidden patterns in massive amounts of data. They are great at sorting patients into "high risk" and "low risk" groups (ranking). However, they are often terrible at telling you the actual probability of an event. They might say, "This patient is very high risk," but when you ask, "What is the chance they survive 5 years?" their answer is often a wild guess.

The Problem: You usually have to choose between the reliable but rigid calculator or the flexible but unreliable super-computer.

The Solution: GRAFT
The authors of this paper created a new model called GRAFT (Gated Residual Accelerated Failure Time). Think of GRAFT as a hybrid vehicle that gets the best of both worlds by splitting the job into two distinct tasks.

1. The "Sorting Hat" vs. The "Weatherman"

Most models try to do two things at once: sort patients by risk and predict the exact timeline. GRAFT says, "Let's stop trying to do both at the same time."

Task A: The Sorting Hat (Ranking). GRAFT uses a powerful, flexible AI brain to look at all the data and figure out who is sicker than whom. It's like a referee sorting a race: "Runner A is faster than Runner B, who is faster than Runner C." It doesn't care about the exact time yet; it just cares about the order.
Task B: The Weatherman (Calibration). Once the Sorting Hat has ranked everyone, GRAFT takes that ranking and runs it through a simple, trusted, old-school formula. This step acts like a "calibration dial." It takes the AI's ranking and translates it into a precise, trustworthy probability (e.g., "There is a 70% chance of surviving 5 years").

By separating these jobs, GRAFT gets the sorting power of AI and the reliability of the old formulas.

2. The "Bouncer" at the Door

Modern medical data is messy. It's like walking into a crowded room where 90% of the people are shouting nonsense, and only a few are giving you useful advice. If you listen to everyone, you'll get confused.

GRAFT has a built-in Bouncer (called a "Stochastic Gate").

As the model learns, this Bouncer walks down the line of data features (like blood pressure, age, genetic markers).
If a feature is just "noise" or irrelevant, the Bouncer gently closes the door and blocks it from entering the model's brain.
If a feature is important, the Bouncer lets it in.

This is crucial because it prevents the model from getting distracted by fake patterns. It learns to ignore the shouting crowd and focus only on the useful voices.

3. The "Best Guess" Strategy for Missing Data

Remember the "censored" patients? We don't know their exact end date.

Old way: Guess a single number for them (e.g., "They probably died 2 years later"). This is risky because if you guess wrong, the whole model learns the wrong lesson.
GRAFT's way: It looks at the patient's neighbors (people with similar symptoms). It says, "Okay, for this group, here is a range of possibilities." Instead of guessing one number, it randomly picks a few different possible end dates from that range and trains on all of them. This teaches the model to be comfortable with uncertainty, making it much more robust.

What Did They Find?

The authors tested GRAFT on six different real-world medical datasets (ranging from breast cancer to HIV).

Better at Sorting: GRAFT was better at ranking patients by risk than almost every other model, including the top AI models.
Better at Predicting: Because of its "Weatherman" step, it was also much more accurate at predicting actual survival times than the AI models.
Noise Proof: When the researchers added a bunch of fake, useless data to the mix (like adding static to a radio signal), the other models got confused and their performance dropped. GRAFT, thanks to its "Bouncer," simply ignored the noise and kept performing well.

The One Catch:
The paper notes that if the data is extremely sparse (meaning almost no one has an event recorded, like in a study where 92% of patients are censored), GRAFT struggles a bit because it needs enough "neighbors" to make its guesses. In those rare, extreme cases, a simpler, old-school model might still be better.

In Summary:
GRAFT is a new tool that splits the job of predicting survival into "sorting" and "calibrating." It uses a smart bouncer to ignore bad data and a flexible strategy to handle missing information. The result is a model that is both flexible enough to handle complex data and reliable enough to trust with real-world decisions.

Technical Summary: GRAFT (Gated Residual Accelerated Failure Time)

Problem Statement

Survival analysis faces a persistent trade-off between interpretability/calibration and discriminative power. Classical models, such as the Cox Proportional Hazards (CoxPH) and parametric Accelerated Failure Time (AFT) models, offer well-calibrated survival estimates and interpretability but are restricted by rigid assumptions (e.g., proportional hazards or specific distributional forms) and linear feature relationships. Conversely, deep learning models (e.g., DeepSurv, DeepHit) capture complex non-linear interactions and achieve strong discrimination but often suffer from poor calibration and sensitivity to high-dimensional noise. Furthermore, real-world clinical datasets are frequently high-dimensional with irrelevant features, yet many models lack robust, integrated mechanisms for automatic feature selection, leading to overfitting and degraded performance in noisy environments.

Methodology: The GRAFT Architecture

The authors propose GRAFT (Gated Residual Accelerated Failure Time), a hybrid model designed to decouple prognostic ranking from survival calibration. The architecture integrates three core components:

1. Hybrid Architecture (Linear + Non-Linear)

Instead of modeling log-survival time directly via a single linear or non-linear function, GRAFT learns a prognostic score $s_i$ that is monotonic with log-survival time. The score is constructed as:
$s_i = \beta^T \phi_i + \mu$
where the representation $\phi_i$ combines:

Gated Linear Component: Input features $x_i$ are masked by a stochastic gate vector $g_{\eta,i}$ (element-wise product), effectively performing feature selection.
Residual MLP: The gated features are passed through a non-linear residual Multi-Layer Perceptron (MLP), $f_\theta(\tilde{x}_i)$ .
This design allows the model to capture both simple linear effects and complex non-linear interactions simultaneously.

2. Stochastic Feature Selection (STG)

To handle high-dimensional noise, GRAFT integrates Stochastic Gates (STG) directly into the architecture.

Mechanism: For each feature $j$ , a gate value $g_{\eta,j}$ is sampled from a continuous relaxation of a Bernoulli distribution (based on a Gaussian distribution with learnable mean $\eta_j$ ).
Training: During training, gates are sampled to create a probabilistic binary mask. A differentiable approximation of the $\ell_0$ norm is added to the loss function to penalize the expected number of active gates, encouraging sparsity.
Inference: At test time, the stochastic sampling is replaced by the deterministic expected value of the gate to ensure reproducible predictions.

3. Training Objective: Decoupled Ranking and Calibration

GRAFT explicitly separates the optimization of ranking (discrimination) from the estimation of survival probabilities (calibration).

Stochastic Conditional Imputation: To handle censored data without assuming a parametric distribution for errors, GRAFT employs local Kaplan-Meier (KM) estimators. For censored subjects, the model constructs a local neighborhood of $k$ nearest neighbors (including censored and event subjects) to fit a conditional KM curve. Event times for censored subjects are imputed by sampling from this conditional distribution (multiple imputation) rather than using a single point estimate.
Differentiable Ranking Loss: The model is trained to optimize the Concordance Index (C-index) by minimizing a differentiable approximation of the negative Spearman rank correlation between predicted scores and imputed log-event times. This uses a soft-ranking operator (based on isotonic regression) to allow gradient-based optimization. Monte Carlo averaging over multiple imputation samples ( $M=5$ ) is used to account for imputation uncertainty.
Post-Training Calibration: The model does not output survival curves directly. Instead, the learned prognostic scores $s_i$ are used as input to a standard CoxPH model (fitted post-training) to estimate the baseline hazard and generate calibrated survival functions $S(t|x)$ . This step ensures calibration without modifying the learned ranking parameters.

Key Contributions

Decoupling of Ranking and Calibration: The paper proposes an explicit separation where the model is trained solely to optimize discrimination via a differentiable ranking objective, while calibration is achieved through a lightweight post-training step. This avoids the trade-offs inherent in single-objective deep learning approaches.
Hybrid AFT Architecture: The model combines a linear AFT framework with a non-linear residual MLP, allowing it to learn both linear and non-linear feature effects while maintaining the interpretability of the AFT structure.
Integrated Stochastic Feature Selection: By embedding Stochastic Gates directly into the architecture, GRAFT performs automatic, differentiable feature selection during training. This provides robustness against irrelevant features and high-dimensional noise, a capability often lacking in standard deep survival models.

Experimental Results

The authors evaluated GRAFT on six public benchmark datasets (GBSG, METABRIC, SUPPORT, NWTCO, FLCHAIN, AIDS) covering varying censoring rates (32% to 92%) and domains (oncology, critical care, HIV).

Performance vs. Baselines: GRAFT consistently outperformed classical baselines (CoxPH, Weibull AFT) and modern deep learning baselines (DeepSurv, DeepHit, Random Survival Forest) in both discrimination (C-index) and calibration (Integrated Brier Score) across most datasets.
- Example: On the SUPPORT dataset, GRAFT achieved a C-index of 0.6143, outperforming DeepSurv (0.6071) and RSF (0.6057).
- High Censoring: In the FLCHAIN dataset (72% censoring), GRAFT achieved the best IBS (0.0919) and a C-index of 0.7965.
Noise Robustness: In ablation studies where noise features were injected (up to 10x the number of original features), GRAFT demonstrated superior robustness. While DeepSurv and DeepHit suffered catastrophic performance degradation, GRAFT maintained stable C-index scores. The ablation study confirmed that the "No STG" variant (removing feature selection) performed significantly worse under noise than the full model, validating the necessity of the gating mechanism.
Limitations: The model showed reduced performance on the AIDS dataset (92% censoring), where the local KM neighborhoods contained insufficient event information for stable imputation. In such extreme censoring scenarios, the authors note that simpler parametric models (like CoxPH) may be preferable.

Significance and Claims

The paper claims that decoupling ranking and calibration is a critical design principle for survival analysis. By optimizing ranking during training and calibrating post-hoc, GRAFT achieves superior performance in both metrics simultaneously, addressing a common failure mode of deep learning survival models.

Furthermore, the authors emphasize that integrated feature selection is essential for robustness in healthcare scenarios characterized by high-dimensional, noisy data. GRAFT's ability to automatically suppress uninformative features allows it to maintain discriminative power where other models degrade. The work suggests that for high-dimensional biomedical applications, a hybrid approach combining the flexibility of deep learning with the structural constraints and calibration properties of classical models offers a practical and effective solution.