A Cox model is the standard tool for predicting time to an event, such as death, when some patients leave the study before the event happens. It assumes that one patient's risk is a fixed multiple of another's at every point in time. When different patient groups have survival curves that cross, that assumption fails, and the usual fix is stratification: give each group its own baseline risk curve. The hard question is which groups to use. We propose taking the groups from a hierarchical community tree over a patient similarity graph. This is a nested set of clusters, from broad to fine, found from network structure. We then prune that tree using survival outcomes alone. Two sibling communities are merged into one stratum, with a simple risk multiplier separating them, whenever the data cannot reject proportional hazards between them. They stay as separate strata only where their hazards genuinely cross. The result is the coarsest set of strata under which a Cox model is valid. It can sit at a different tree depth in different branches, and it keeps the interpretability of hazard ratios inside each stratum.
The opportunity is that both fields have half of the answer. Flat latent mixtures such as Deep Cox Mixtures and Survival Cluster Analysis show that splitting a population into groups with separate baselines repairs calibration when hazards cross. Both, however, choose one flat resolution and ignore any network among patients. Multiscale community methods, such as the Markov multiscale clustering used by Beaney et al. on 10.5 million primary-care records, find nested disease groupings, but they never touch outcomes. Those authors name moving from disease clusters to patient clusters with outcomes as open work. Our claim is narrow and testable. When a comorbidity graph is available from a different data source than the Cox covariates, a bottom-up community tree pruned by a proportionality test gives better-calibrated survival predictions than flat mixtures and covariate-split survival trees, using no more strata than they do. The method also reports which branches of the hierarchy break proportional hazards.
Core Claim
The unit of novelty is the pruning criterion, not the tree and not the Cox model. Let $T$ be a dendrogram of patient communities, and let a cut $C$ be any set of tree nodes that partitions the patients. For a cut $C$, each stratum $s$ is a tree node. Each finer community $c$ beneath it, at the level where we stop pruning, receives a time-constant offset $\gamma_c$. The model is fit by the stratified partial likelihood
$$L(\beta,\gamma; C) = \prod_{s \in C} \prod_{i \in s:\ \delta_i = 1} \frac{\exp(\beta^\top x_i + \gamma_{c(i)})}{\sum_{j \in R_s(t_i)} \exp(\beta^\top x_j + \gamma_{c(j)})},$$
where $x_i$ are patient covariates, $\delta_i$ indicates an observed event, $c(i)$ is patient $i$'s community, and $R_s(t_i)$ is the set of patients in stratum $s$ still at risk at time $t_i$.
At each internal node $v$ we test whether its children can share a stratum. We let each child's offset drift in time, $\gamma_c(t) = \gamma_c + \eta_c \log t$, and test $\eta_c = 0$ with a likelihood-ratio test. If the test does not reject, the children merge into a stratum with constant offsets. If it rejects, they stay as separate strata. Pruning proceeds bottom-up. To avoid testing on the same data that built the tree, the tree is built on one fold and the tests are run on another. Multiplicity across nodes is controlled by testing parents only after their children, with a hierarchical error-rate procedure in the spirit of the main-effects-first ordering of Liang et al.. Their paper applies that ordering to covariate interactions in an accelerated-failure-time model; we borrow only the testing order.
Research Program
The tree comes from the method of Dreveton et al.. That method finds fine "primitive" communities using the Bethe-Hessian matrix, a spectral tool built for sparse graphs. It then merges them bottom-up by average linkage. Their analysis shows that this ordering avoids dendrogram inversions, where a child cluster is less cohesive than its parent, which top-down splitting suffers from. It also recovers communities down to near-constant average degree. Their own caveat, that the whole tree inherits errors from the leaf partition, constrains our design. We therefore also run the pruning on Bayesian Hierarchical Community Discovery trees. That method builds non-binary "rose trees" by greedy Bayesian merging, which gives a second, independently built hierarchy and so measures how sensitive the strata are to the tree builder.
Separating the graph from the covariates is the design choice that makes the graph matter. In MIMIC-III, which Tjandra et al. used for multi-event survival, we build the patient graph from shared diagnosis codes, using Jaccard-weighted k-nearest-neighbour links. The Cox covariates are demographics, vitals and laboratory values only. Comorbidity communities can then define baseline-risk strata that a covariate-only model cannot see.
This setup also addresses the tension the research file raises between assortative graphs and crossing risks. The graph is assortative by construction: similar patients are linked. Nothing assumes hazards are. The proportionality test decides, branch by branch, whether network closeness implies proportional risk.
The analysis yields three things. The first is a selected cut, whose depth can differ across branches. The second is a per-node record of where proportionality fails, which answers the question in the research file's Gap 2 of where in the hierarchy proportional hazards breaks down. The third is calibrated individual survival curves from Breslow-type baselines in each stratum.
Baselines
The naive reference is a single-stratum linear Cox model on the same covariates. Deep Cox Mixtures reports an expected calibration error (ECE) of 0.0992 for it on FLCHAIN. DeepSurv is the nonlinear single-baseline reference. It reports a METABRIC C-index of 0.643 against 0.630 for linear Cox, and the BMC version reports 0.867 against 0.816. This discrepancy is why we rerun every model ourselves rather than quote numbers.
The strongest existing competitor is Deep Cox Mixtures. It learns $K$ latent groups, each with its own baseline and neural hazard network, by hard-assignment expectation-maximization. It is the method that already fixes crossing-hazard miscalibration, and it reports FLCHAIN ECE 0.0200. We run it with $K$ tuned over the same range as our number of selected strata, plus once with $K$ fixed to our selected count. Both arms get the same covariates, splits, and hyperparameter-search budget. Beating it means lower ECE and integrated Brier score at a C-index within noise, with no more strata. On covariate-only FLCHAIN, where our graph is merely a k-nearest-neighbour graph on covariates, we expect at best a tie. We report FLCHAIN as a test of whether our tree-and-prune procedure gives up anything relative to flat mixtures, not as the venue for winning. For MIMIC-III and for DCM on SUPPORT, we found no reported result on Lacuna.
Survival Cluster Analysis is run as a second flat-mixture baseline. It reports a SEER log-rank score of 4803.59 and an FLCHAIN calibration slope of 0.9900.
The ablations isolate the key ingredient. The first uses the same tree cut at the single fixed level with the best validation likelihood. The second uses the same tree pruned by C-index gain instead of the proportionality test. The third uses the same pruning with the graph edges randomly rewired while degrees are preserved. The rewiring arm is decisive: if it matches the full method, the graph contributes nothing and the proposal fails.
Why not just…?
Why not just fit Deep Cox Mixtures and pick $K$ by validation? This is the most serious threat. DCM's mixture assignments are driven by covariates, so if the comorbidity information is already mostly captured by labs and vitals, DCM will match us. The flat structure also cannot express "broad group with a shared baseline except one sub-branch that crosses". Such a pattern would force DCM to spend extra components. We settle this with the matched-$K$ comparison above, and with a variant that feeds DCM's gating network the diagnosis codes directly. That variant gives a flat model the same information our graph has, so any remaining gap is attributable to hierarchy plus outcome-based pruning rather than to extra inputs.
Why not just use a survival tree, which is already a hierarchy with per-leaf baselines? Survival trees split greedily from the top, on covariate thresholds, and optimize separation of outcomes rather than validity of proportional hazards. On planted subgroups, Izzo and Melvin report F1 0.28 for survival trees and 0.97 for their method, which finds axis-aligned sub-cohorts where the Cox model holds tightly. We include both as baselines. Their method is the closest competitor in spirit, but it works only on axis-aligned regions of covariate space and cannot use graph structure that lives in a separate modality. If either matches us on MIMIC-III, our graph-derived hierarchy is unnecessary.
Why not just cluster the graph once and add community labels as Cox covariates? Labels entered as covariates still impose proportional hazards on every community. Labels used as strata at a fixed cut ignore outcomes when choosing resolution. Survival Cluster Analysis argues that outcome-blind clustering yields groups that differ in features but not in risk. The fixed-level-cut ablation is exactly this pipeline, so the experiment measures the value of outcome-based pruning directly.
Proposed Experiments
The first study is synthetic. We generate patients from a hierarchical stochastic block model, a random-graph model with planted nested communities. The planted strata sit at mixed depths: some branches have crossing baseline hazards at a deep level, while others are proportional all the way up. We vary graph sparsity and censoring rate. We score recovery of the true stratum partition with F1 and adjusted Rand index, and score calibration with ECE. This tests whether the proportionality-based pruning finds mixed-depth cuts that no single-level cut or flat $K$ can represent. It also probes the leaf-error sensitivity that Dreveton et al. warn about, by degrading the graph until primitives are misrecovered.
The main study uses MIMIC-III in-hospital and one-year mortality, with the comorbidity graph described above. Metrics are time-dependent C-index, integrated Brier score, ECE, and number of strata. We use five cross-fitted splits and report paired confidence intervals. FLCHAIN and SUPPORT serve as covariate-only checks, comparing against DCM's reported FLCHAIN ECE of 0.0200 and our own reruns.
Two checks decide the proposal. The first: on MIMIC-III, the full method must beat both matched-$K$ DCM given diagnosis codes and the method of Izzo and Melvin on ECE and integrated Brier score without losing C-index. The second: the degree-preserving rewired-graph ablation must lose most of that gain. If the first fails, hierarchy plus pruning adds nothing over flat mixtures. If the second fails, the network is decoration, and the honest conclusion is that covariate-space subgroup discovery suffices.
References
- Deep Cox Mixtures for Survival Regression
- Survival Cluster Analysis
- Identifying multi-resolution clusters of diseases in ten million patients with multimorbidity in primary care in England
- Hierarchical false discovery rate control for high-dimensional survival analysis with interactions
- When Does Bottom-Up Beat Top-Down in Hierarchical Community Detection?
- Bayesian Hierarchical Community Discovery
- A Hierarchical Approach to Multi-Event Survival Analysis
- Deep Survival: A Deep Cox Proportional Hazards Network
- DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network
- Subgroup Discovery with the Cox Model