Pruning Random Survival Forests with a Censoring-Correct Error Covariance

Random survival forests average hundreds of decision trees to predict, for each patient or customer, a survival curve: the probability of still being event-free at each future time. In classification, it is well established that a carefully chosen fraction of a bagged forest can beat the whole forest. Two examples: discarding up to 85% of bagging members can raise accuracy (Pruning in ordered bagging ensembles), and a 4-of-10 deep sub-ensemble can outperform the full set (Hierarchical Pruning of Deep Ensembles with Focal Diversity). Nobody has done this for survival forests. The reason is not lack of interest. Pruning criteria need a per-example notion of "this tree was wrong here," and in survival data many subjects are right-censored: they left the study before the event, so we only know the event happens after their last observation. For a censored subject, "wrong" is undefined. Treating them as events biases results toward early failure, and dropping them biases results toward long survival.

This note proposes a simple remedy. A standard censoring-aware accuracy measure for survival curves is the integrated Brier score (IBS): squared error between predicted survival probability and observed survival status, averaged over time, with each subject reweighted by the inverse probability of remaining uncensored. The IBS of a uniformly averaged sub-forest is exactly a quadratic function of which trees are included. That quadratic has a clean reading. Its diagonal entries are each tree's own accuracy. Its off-diagonal entries are a censoring-corrected covariance of the trees' errors, which gives a principled "diversity" measure for censored data. Once that matrix is available, tree pruning for survival reduces to a well-studied selection problem. Survival forests are known to overfit on small cohorts and to drown clinical signal in molecular noise, so they are the natural place to test whether pruning helps.

Core Claim

For each tree $m$, write $S_m(t\mid x_i)$ for its predicted survival probability for subject $i$ at time $t$. Write $y_i(t)=1\{T_i>t\}$ for the observed survival status. Write $w_i(t)$ for the standard inverse-probability-of-censoring weight: $1/G(t)$ if the subject is still under observation at $t$, $\delta_i/G(T_i)$ if the event was observed before $t$, and $0$ if the subject was censored before $t$. Here $G$ is the Kaplan-Meier estimate of the censoring distribution. Define the censoring-weighted error Gram matrix over the time window $[0,\tau]$:

$$\Gamma_{jk}=\frac{1}{n\tau}\sum_{i=1}^{n}\int_0^{\tau} w_i(t)\,\big(S_j(t\mid x_i)-y_i(t)\big)\big(S_k(t\mid x_i)-y_i(t)\big)\,dt .$$

Then for any subset $A$ of $K$ trees whose survival curves are averaged uniformly, the IBS is $\mathrm{IBS}(A)=\frac{1}{K^2}\sum_{j,k\in A}\Gamma_{jk}$. This holds as an identity, because the IBS is a weighted squared error and the ensemble error is the mean of the member errors.

The claim is that minimizing this quadratic over subsets yields sub-forests that are smaller and better calibrated than the full forest. The minimization uses ordered greedy aggregation, with a semidefinite relaxation as a check. The matrix $\Gamma$ is estimated from out-of-bag or cross-fitted predictions. Crucially, the claim is that using the censoring weights is what makes this work: naive error matrices that mishandle censored subjects should fail, and fail increasingly as censoring grows.

The contribution is therefore not "apply ensemble pruning to survival." It is a specific object, $\Gamma$, that makes existing pruning machinery statistically correct under right-censoring. The proposal also includes a direct test of whether that correctness matters.

Relation to Prior Work

The closest work is Experimental Comparison of Ensemble Methods and Time-to-Event Analysis Models Through Integrated Brier Score and Concordance Index. It combines six heterogeneous survival models (Cox, Weibull, Aalen, random survival forest, gradient-boosted Cox, DeepSurv) with continuous simplex weights tuned by exponential gradient descent on IBS. That paper supplies the objective and motivates the work: it shows IBS-driven combination helps. It differs from this proposal in three ways. It keeps all models, it never selects a discrete subset, and its pool is six models. With six models, exhaustive search over 64 subsets is trivial, so pruning there would be uninteresting. We work inside a forest of hundreds of trees on cohorts as small as 276 subjects. In that regime, fitting one continuous weight per tree is an overfitting risk rather than a free improvement.

Three pruning methods supply the machinery.

  • Topology-informed pruning (Ensemble Pruning for Out-of-distribution Generalization) solves a quadratic integer program over a Gram error matrix by semidefinite relaxation. We reuse that solver, but our Gram matrix is censoring-weighted and IBS-exact, and we choose the sub-ensemble size from the objective rather than fixing it at half the pool.
  • Focal-diversity hierarchical pruning (Hierarchical Pruning of Deep Ensembles with Focal Diversity) measures how well models cover each other's failures. The off-diagonal of $\Gamma$ is a censored-data analogue that needs no binary notion of failure.
  • Pareto pruning (Pareto Ensemble Pruning) trades error against size. We use its frontier view to report IBS against the number of trees.

The HTrees permutation test (A Robust Hypothesis Test for Tree Ensemble Pruning) prunes splits within gradient-boosted trees and explicitly does not handle censored targets. We prune whole trees, so we avoid the problem of permuting $(T,\delta)$ pairs.

Evidence that forests need this comes from two sources.

Both results suggest that many trees add noise rather than signal. X-CAL: Explicit Calibration for Survival Analysis constrains the proposal. It shows that ranking quality and calibration diverge. We must therefore report concordance and D-calibration alongside IBS, and not claim that an IBS gain implies both.

Research Program

We grow random survival forests of 500 to 1000 trees. Each tree is fit on a subsample without replacement at rate one half, rather than a bootstrap. With the usual bootstrap, only about 13% of subjects are out-of-bag for both trees in a pair. With half-subsampling, about 25% are, which gives enough overlap to estimate $\Gamma_{jk}$ on a 276-subject cohort. As a second estimator, we also compute $\Gamma$ from K-fold cross-fitted forests.

The forest aggregates survival curves by averaging them, which is the operation the identity requires. The usual random survival forest instead averages cumulative hazards, so we first check that this change does not hurt the full forest.

Selection uses ordered aggregation. Starting from the best single tree, we repeatedly add the tree that most reduces the quadratic. This produces a full path of sub-forest sizes, and we pick $K$ at the minimum of a shrinkage-regularized estimate of IBS along that path. A semidefinite relaxation of the fixed-$K$ problem checks how far greedy is from the relaxed optimum.

The analysis has three parts:

  • how the chosen $K$ and the IBS gain scale with censoring rate and sample size;
  • whether the retained trees in multi-omics data preferentially split on clinical versus molecular features;
  • a secondary application of the same $\Gamma$ to a large heterogeneous pool, namely many hyperparameter settings of Cox, Weibull, DeepSurv, and multi-task logistic regression, to test generality beyond trees.

Baselines

The naive reference is the full forest of all trees, averaged uniformly, with the same trees, splits, and seeds as the pruned version. Any pruning claim must beat it on held-out IBS at a fraction of the trees.

The tuned small forest is a random survival forest whose number of trees, features per split, and minimum node size are tuned by nested cross-validation. It gets the same number of trees as the pruned sub-forest and a matched tuning budget. This tests whether pruning is just a roundabout way to regularize.

The strongest existing method is the convex exponential-gradient ensemble of Experimental Comparison of Ensemble Methods and Time-to-Event Analysis Models Through Integrated Brier Score and Concordance Index. We run it in two ways. First, over its original six model types, using the same splits. Second, over the individual trees of our forest, with weights fit on the same out-of-bag data used to estimate $\Gamma$. That paper reports a 3% IBS improvement over DeepSurv on PBC and a 3.4% improvement over the "average best" single model across PBC, GBCSG2, and TLCM. The file gives no absolute IBS values, so we will reproduce them under our protocol rather than quote them. Beating this baseline means lower held-out IBS than its tree-level variant with no loss in concordance. If we instead match its six-model result with a single pruned forest, that is also a meaningful outcome.

For discrimination context on METABRIC, DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network reports concordance of 0.867 for DeepSurv, 0.816 for Cox, and 0.619 for random survival forest. A pruned forest should move clearly above 0.619 without dropping calibration. We do not expect pruning alone to reach 0.867.

On TCGA we add a clinical-only Cox model and Block Forest. No per-cohort reported result was found on Lacuna for TCGA, so we will run both ourselves.

The key ablation keeps the pruning algorithm identical but replaces $\Gamma$ with error matrices that mishandle censoring. There are three variants:

  • setting all weights $w_i(t)$ to one, which treats censored subjects as survivors forever;
  • dropping censored subjects;
  • using per-tree concordance plus prediction correlation, a focal-diversity-style heuristic.

Why not just…?

Why not just use the continuous IBS-weighted ensemble instead of pruning? This is the real threat. Minimizing a convex quadratic over the simplex often yields sparse weights on its own, so thresholding exponential-gradient weights might reproduce our subsets. The file has no evidence either way at the scale of individual trees. The tree-level exponential-gradient baseline, with and without thresholding, settles the question directly. If it matches uniform pruning, we will report that the contribution reduces to the censoring-correct Gram matrix plus sparsity of that problem. We expect uniform subsets to win on the smallest cohorts, where fitting hundreds of weights on around 276 subjects should overfit.

Why not just tune a smaller random survival forest? Tuning changes how trees are grown. Pruning selects among grown trees using their joint errors, which tuning cannot see. The matched-size tuned-forest baseline tests this. If it matches pruning, the pruning result is not interesting and we will say so.

Why not just calibrate a single model with X-CAL or use DeepSurv? X-CAL: Explicit Calibration for Survival Analysis cuts D-calibration error, but only within one parametric neural model and with added tuning. Meanwhile, Experimental Comparison of Semi-parametric, Parametric, and Machine Learning Methods for Time-to-Event Analysis Through the IPEC Score finds that the best single model changes by dataset: Cox wins on TLCM, where DeepSurv ranks fifth, and DeepSurv wins on GBCSG2. We include DeepSurv, with and without the X-CAL penalty, as reference points. Our claim concerns improving forests, not beating every model family.

Why not just use Block Forest for multi-omics? Block Forest requires predefined feature groups, which do not exist for TLCM or SUPPORT, and its gain on TCGA was marginal. We compare against it on TCGA, where groups exist, and also test whether pruning a Block Forest improves it further.

Proposed Experiments

Datasets.

  • Low-dimensional clinical and customer cohorts: PBC, GBCSG2, and TLCM, as in the convex-ensemble study.
  • METABRIC, SUPPORT, and WHAS, as in the DeepSurv study.
  • The high-dimensional DBCD and DLBCL sets from the individual-survival-distribution suite.
  • The 18 TCGA multi-omics cohorts.
  • Synthetic censoring added to PBC and TLCM at 20%, 40%, 60%, and 80%.

Protocol. We use repeated 5-fold outer cross-validation, with all selection done inside the training folds. We report held-out IBS, Uno's concordance, D-calibration, and the fraction of trees retained.

Decisive checks. There are three.

  • On at least most of the low-dimensional cohorts, a sub-forest of at most a quarter of the trees achieves lower held-out IBS than the full forest, with concordance within fold-to-fold noise.
  • The advantage of the censoring-correct $\Gamma$ over the unweighted and censored-dropped ablations grows monotonically with the synthetic censoring rate. If it does not, the core claim fails.
  • On TCGA, pruned forests narrow the gap to the clinical Cox model relative to full forests, and the retained trees are enriched for clinical splits.

A negative result on the first check alongside a positive result on the second would still be informative: it would show that censoring-correct diversity matters even where pruning itself gives little gain.