Doctors and reliability engineers who predict when something will happen usually face a forced choice. They can use a Cox model, a regression that estimates how each risk factor multiplies the instantaneous risk of the event (the "hazard") and that copes with patients who leave a study early. Or they can use a neural network, which predicts better but cannot say why. The Cox model's clarity also depends on two assumptions that often fail. It assumes each factor acts as a straight line on the log-risk scale, and it assumes each factor's effect stays the same over the whole follow-up period (the "proportional hazards" assumption). This note proposes a model that is transparent by construction, so that inspecting it means reading its terms, and that drops both assumptions while still being trained on the Cox model's own objective.
The model is a sum of one-variable curves, a few two-variable surfaces, and a small budget of variable-by-time surfaces. Each surface shows directly how a risk factor's hazard ratio drifts over follow-up. All of these pieces are fit by gradient boosting on the Cox partial likelihood, which is the objective that compares each patient who has the event against everyone still at risk at that moment. The exact gradients and curvatures come from the linear-time identities in FastSurvival. The practical promise is a clinician-readable hazard model that keeps censoring and continuous time intact. It avoids the data blow-up that current interpretable pipelines incur, and it shows non-proportional effects as pictures rather than hiding them.
Core Claim
Three systems in the file bracket the gap, and each gives up something different. DeepSurv replaces the Cox linear score with a neural network trained on the partial likelihood. It reaches a concordance index (C-index, the probability that the model ranks a random pair of patients in the correct order of event time) of 0.643 on METABRIC, against 0.630 for linear Cox, but its authors concede the model is a black box. FastSurvival supplies the method we reuse. It shows that the partial-likelihood derivatives are the mean and variance of covariates under a softmax distribution over each risk set, and that reverse cumulative sums compute them in $O(n)$ for $n$ patients. Its model, however, is strictly linear and time-invariant.
Interpretable Survival Analysis for Heart Failure Risk Prediction is the closest prior work. It reaches non-linear interpretability with Explainable Boosting Machines (EBMs, additive models with curves and pairwise surfaces). To get there, it converts survival into binary classification through "survival stacking": the dataset is copied once per event time, which multiplies its size by up to the number of event times $K$. Even after undersampling, the authors report memory and compute bottlenecks on their 363,398-patient cohort. That paper also motivates our work, because it found clinically meaningful non-linear age-by-EKG interactions that a linear Cox model misses.
Our claim is that additive curves, pairwise surfaces, and explicitly time-varying curves can all be fit on the continuous-time partial likelihood. The cost is $O(nB)$ per boosting round, where $B$ is a small number of time bins, instead of the $O(nK)$ data expansion of stacking. The result should match stacked EBMs and DeepSurv in discrimination, beat proportional-hazards models whenever effects change over time, and turn proportional-hazards violations into readable time-by-variable plots.
Model and Fitting
The hazard of patient $i$ with covariates $x$ at time $t$ is
$$h(t \mid x) = h_0(t)\exp\Big(\sum_k f_k(x_k) + \sum_{(k,l)\in S} f_{kl}(x_k,x_l) + \sum_{k\in T} g_k(x_k)\,\mathbb{1}[t \in \tau_{b(t)}]\Big),$$
where the terms are defined as follows. The baseline hazard $h_0(t)$ is left unspecified. Each $f_k$ is a one-variable shape curve. The set $S$ holds a few selected variable pairs with surfaces $f_{kl}$. The set $T$ holds variables allowed to have time-varying effects. The function $g_{k}$ is piecewise constant over $B$ time intervals $\tau_1,\dots,\tau_B$, and $b(t)$ indexes the interval containing $t$. In practice $g_k$ is a two-dimensional surface over (value of $x_k$, time bin). Its slice at a given bin is the extra log hazard ratio contributed by $x_k$ during that period of follow-up. If the slices are flat across bins, proportional hazards holds for that variable; if they tilt, it fails, and the plot shows how.
Fitting uses cyclic gradient boosting in the EBM style: shallow trees on one term at a time, with a small learning rate. Each leaf takes a Newton step. The step uses the leaf's sum of first derivatives divided by its sum of second derivatives of the negative log partial likelihood.
Following FastSurvival, these derivatives for all $n$ patients come from reverse cumulative sums of $\exp(\text{score})$ over patients sorted by time. Time-varying terms make a patient's score depend on which bin the current event time falls in. Within one bin, however, every score is fixed. Running the cumulative sums bin by bin, with each bin's scores, therefore gives exact derivatives in $O(nB)$ with no dataset copies. Ties are handled with the Breslow approximation.
Pairs for $S$ and $T$ are chosen by the usual FAST-style screening, now scored by partial-likelihood gain. The candidate pairs for $T$ are variable-by-time, and they get their own budget, so that time-variation is admitted only where the data support it.
Correlated features cause importance to be split arbitrarily among near-duplicates. Two corpus papers flag this problem: the heart-failure paper, which prunes with ControlBurn, and FastSurvival, which shows Newton solvers break down when features are correlated at $\rho=0.9$. To address it, we add a final sparsity stage over whole terms. Each fitted term's contribution is treated as one derived feature, and FastSurvival's $L_0$-constrained coordinate descent is run over these term-features. This chooses which terms survive.
Research Program
The work has three parts. The first is an open-source implementation with exact binned derivatives, Newton leaf steps, interaction screening by partial-likelihood gain, and the term-level $L_0$ pruning stage.
The second is a synthetic study where the truth is known. We build hazard functions with U-shaped effects, crossing survival curves, delayed treatment effects, and clusters of near-duplicate features. We then measure whether each method recovers the true curves and the true time profiles, not merely whether it ranks patients well.
The third is a benchmark comparison on the public survival datasets in the file. If access can be arranged, we also test on a large EHR cohort comparable to the heart-failure study, to stress memory against stacking. A secondary analysis asks whether time-varying surfaces explain gains that DeepSurv attributes to non-linearity. If a deep model's advantage disappears once time-variation is allowed, then some of what looked like complex interaction was really a proportional-hazards violation.
Baselines
The naive reference is linear Cox with $L_1$ or $L_0$ sparsity, fit with FastSurvival's solver. It is the "do nothing new" option, because it is already interpretable and handles correlated features with 100% support recovery (15 of 1,200 variables at $\rho=0.9$) in FastSurvival. Beating it means a clear C-index gain on real data, plus recovery of non-linear and time-varying truth on synthetic data that linear Cox cannot represent.
The strongest discrimination competitor is DeepSurv. On METABRIC its reported C-index is 0.643, versus 0.630 for linear Cox and 0.624 for random survival forests. We will rerun it rather than copy these figures.
The closest interpretable competitor is the stacked-EBM pipeline from the heart-failure paper. For this pipeline we found no reported result on Lacuna on METABRIC, GBSG, SUPPORT, WHAS, Flchain or the other public sets. The paper reports only AUC comparisons, with no exact values in the file, and only on a private EHR cohort, so we will reproduce it ourselves.
The strongest classical competitor is a penalized-spline additive Cox model with spline-by-time interactions. This is the textbook statistical answer and is discussed further below.
All methods get identical repeated five-fold splits, the same hyperparameter search budget (a fixed number of random-search trials per method, chosen by inner-fold validation partial likelihood), and the same hardware. Wall-clock time and peak memory are reported alongside accuracy.
Our success criterion has two parts. On METABRIC, we must land within the fold-to-fold standard error of the DeepSurv reproduction. On the synthetic non-proportional data and at least two real datasets with detected proportional-hazards violations, we must beat stacked EBMs and spline Cox on time-dependent concordance and integrated Brier score (the time-averaged squared error of predicted survival probabilities).
Three ablations isolate the key ingredients. The first removes all time-varying terms (a proportional-hazards additive model). The second replaces Newton leaf steps with gradient-only steps. The third trains the same booster on stacked data instead of the partial likelihood, separating the effect of the objective from the effect of the model class.
Why not just…?
Why not just fit a penalized-spline additive Cox model with time-varying coefficients? This is the most serious threat, and on small datasets with a handful of covariates we expect it to tie. Where it should fall short is choosing which of hundreds of correlated features get curves, pairs, or time terms. Spline Cox needs those choices specified in advance, and its Newton solver faces the same collinearity failures FastSurvival documents. We will include it on every dataset where it can be fit. If it matches us everywhere, including the synthetic correlated-feature setting, the contribution shrinks to engineering and we will say so.
Why not just train DeepSurv and explain it with SHAP? Post-hoc attributions can be ambiguous even when the model is confident, as shown by To believe or not to believe, where LIME heatmaps over API sequences stayed unclear at roughly 95% accuracy. Attributions also split across correlated variables, which is the problem the heart-failure paper had to prune around. On synthetic data we will measure directly how far SHAP curves from DeepSurv deviate from the true shape functions, compared with our fitted terms. Most proportional-hazards attributions also cannot express a time-varying effect at all.
Why not just use the stacked EBM pipeline, perhaps with mini-batching? Its authors suggest mini-batching but did not implement it. Even so, stacking optimizes a discretized binary surrogate rather than the partial likelihood and still touches $O(nK)$ rows. The objective ablation, the same booster on stacked versus partial-likelihood data, will show whether the continuous objective matters for accuracy or only for memory. Either outcome is informative.
Why not just predict a fixed-horizon binary outcome with an EBM? This drops or mislabels censored patients and discards event timing. It is included implicitly as a 5-year-horizon EBM reference on GBSG and METABRIC.
Proposed Experiments
The synthetic suite crosses three factors: shape (linear, U-shaped, interacting), time behavior (constant, delayed, crossing), and correlation (none, or $\rho=0.9$ blocks following FastSurvival's 1,200-feature design). Censoring rates range from 20% to 70%. The metrics are integrated squared error between fitted and true log-hazard-ratio terms over time, term-level support recovery, and time-dependent C-index.
The real-data suite uses METABRIC, GBSG with Rotterdam as an external test set, SUPPORT, WHAS, Flchain, Kickstarter and EmployeeAttrition. The metrics are Harrell's C-index, time-dependent concordance, integrated Brier score, and calibration at fixed horizons. Proportional-hazards violations are first flagged with Schoenfeld residual tests (which check whether a covariate's effect drifts with time) on a linear Cox fit, so that improvements can be tied to variables known to violate the assumption.
A scaling experiment subsamples a large cohort from $10^4$ to $3\times 10^5$ patients and records memory and time for our method against stacked EBMs. We expect our cost to grow linearly in $n$ for fixed $B$, while stacking grows with $n$ times the number of distinct event times.
Three results would be decisive. Showing that time-varying surfaces recover crossing-hazard truth where every proportional-hazards model, deep or not, cannot would confirm the method's distinctive value. Showing that the DeepSurv gap on METABRIC closes once time terms are added would locate where black-box gains actually come from. A stacking-versus-partial-likelihood ablation with no accuracy difference would reduce the claim to efficiency, and we will report it that way.