Non Parametric Tree Ensembles for Survival Analysis

80 concepts57 papers1 proposal

In many real-world applications, we do not just want to know if an event will happen, but when it will happen. This is the domain of survival analysis (or time-to-event modeling). It is used to predict when a heavy-duty truck battery will fail (Frisk et al.), how long a user will stay on an advertisement's landing page (Barbieri et al.), or the life expectancy of a patient diagnosed with cancer (Ingrisch et al.).

Traditional statistical methods, most notably the Cox Proportional Hazards model, have long been the default tools for this task. However, they rely on rigid assumptions—such as requiring that the relative risk between two individuals remains constant over time, and that features interact linearly.

Non-parametric tree-based ensembles, such as Random Survival Forests (RSF) and boosted survival trees, fundamentally change this approach. By adapting the core mechanics of decision trees to handle the unique geometry of time-to-event data, these models capture complex, non-linear interactions automatically without requiring predefined parametric distributions.

The Challenge of Right-Censoring

To understand why standard machine learning regression fails at survival analysis, we must look at right-censoring.

Imagine a clinical trial tracking patient survival over five years. If a patient drops out of the study in year two, or if the study ends while they are still alive, we do not know their exact time of death. We only know they survived for at least two or five years. Dropping these individuals introduces severe bias, while treating their censored times as their actual event times underestimates true survival.

Standard regression models minimize a loss function based on exact targets (like Mean Squared Error). Survival models, instead, must predict a continuous function: the Survival Function $S(t)$, which represents the probability of surviving past time $t$.

How a Tree Learns to Split Survival Data

A standard decision tree splits data by minimizing variance (for regression) or Gini impurity (for classification). To handle censored time-to-event data, survival trees modify two fundamental components: the splitting criterion and the leaf node prediction.

1. The Splitting Criterion: Maximizing Survival Divergence

Instead of looking at the mean value of a target variable, a survival tree evaluates a split by determining how well it separates individuals into groups with distinctly different survival trajectories.

The most common metric used is the log-rank test statistic. At each candidate split, the algorithm compares the survival experiences of the proposed left and right child nodes. It calculates the observed versus expected number of events at every distinct time point across both groups.

A high log-rank statistic indicates that the two groups have highly divergent survival curves. By maximizing this statistic, the tree isolates high-risk individuals from low-risk individuals, naturally uncovering critical thresholds—such as specific lymph node counts in cancer staging (Yuan et al.).

2. The Leaf Node Prediction: Estimating the Curve

In a standard tree, a leaf node outputs the average value or majority class of the training samples that land inside it. In a survival tree, a leaf node outputs an entire non-parametric survival curve.

Once the tree stops splitting, the algorithm takes all the samples inside a given leaf node—both censored and uncensored—and computes a step-function estimate of survival, typically using the Kaplan-Meier or Nelson-Aalen estimator.

$$\hat{S}(t) = \prod_{t_i \le t} \left(1 - \frac{d_i}{n_i}\right)$$

Where $d_i$ is the number of events at time $t_i$, and $n_i$ is the number of individuals still at risk just before $t_i$. Because this calculation happens locally within the leaf, the model makes no global assumptions about the shape of the risk over time.

[Root Node: All Patients]
          |
          v  (Split maximized by Log-Rank Test)
     +----+----+
     |         |
     v         v
[Leaf A]    [Leaf B]
  |           |
  v           v
[Kaplan-Meier Curve A]  [Kaplan-Meier Curve B]
(Low Risk Profile)      (High Risk Profile)

Why Ensembles Excel: Smoothing and Interaction Capture

While a single survival tree provides a highly interpretable set of rules, its predicted survival curves are jagged step-functions that are prone to overfitting. Ensembling multiple trees solves this.

Bagging and Boosting

In a Random Survival Forest, hundreds of survival trees are trained on bootstrapped samples of the data. To predict the survival curve for a new individual, their features are passed down every tree in the forest. Each tree deposits the individual into a specific leaf node and returns that leaf's unique Nelson-Aalen or Kaplan-Meier curve. The forest then averages these curves together.

This averaging process smooths out the step functions into a continuous, highly calibrated individual survival distribution. Alternatively, boosting frameworks can be applied to survival trees. As Alexis Bellot et al. demonstrate in Boosted Trees for Risk Prognosis, combining shallow survival trees sequentially via gradient boosting can yield superior risk discrimination and probability calibration by learning the shape of the survival curve directly from data.

Automatic Discovery of Non-Linearities

Because tree ensembles recursively partition the feature space, they excel at discovering complex biological or behavioral interactions without manual feature engineering.

For instance, in life insurance underwriting, risk factors often exhibit U-shaped relationships (where both exceptionally low and exceptionally high values of a biomarker indicate elevated mortality risk). Marc E. Maier et al. showed that non-parametric survival forests naturally map these U-shaped risks, outperforming traditional linear underwriting models (the authors).

Similarly, in precision medicine, tree ensembles can seamlessly combine high-dimensional radiomic imaging features (Ingrisch et al.) or genetic variations (Zhang et al.) with standard clinical data, capturing multi-dimensional interactions that violate the proportional hazards assumption of classical models.

Advanced Variations

The flexibility of tree-based survival architectures allows them to be extended to handle complex data structures that traditional regression models struggle to accommodate.

Competing Risks

In many studies, an individual can experience multiple types of distinct events that prevent the primary event of interest from occurring (e.g., a patient dying from a heart attack before their cancer recurs).

To address this, Alexis Bellot et al. developed a multitask boosting framework (Multitask Boosting for Survival Analysis with Competing Risks) that embeds the Aalen-Johansen estimator into the leaf nodes of decision trees. This allows the ensemble to estimate the cumulative incidence function for multiple competing risks simultaneously, borrowing statistical strength across different event types without relying on rigid distributional assumptions.

Clustered Data and Shared Frailty

When survival data is collected from clustered groups (such as multiple patients treated at the same hospital, or repeated observations of the same asset), the assumption of independence between observations breaks down.

As David M. Zucker reviews, tree-based algorithms can be adapted for clustered data by incorporating random effects—or frailty terms—directly into the node-splitting criteria (the author). By modifying the splitting logic to account for group-level variances, the model can successfully distinguish between true covariate effects and shared unobserved environmental factors.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_random_survival_forests_a5265990,
  title = {random survival forests},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/random-survival-forests/rdr_a52659904a1c4a5cb9e7db6fdf35294c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_random_survival_forests_a5265990_v1,
  title = {random survival forests},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/random-survival-forests/rdr_a52659904a1c4a5cb9e7db6fdf35294c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake