The White-Box Survival Model on Trial

A kernel machine is one of the most beautiful objects in machine learning, and its dense textbook implementation can be inconvenient. A fixed RKHS comes with a norm and, for the right regularized objectives, a representer theorem that reduces the optimizer to a finite expansion over training inputs. Dense support-vector and kernel-ridge baselines can require quadratic storage or expensive solves, but that is not the whole algorithmic landscape: stochastic optimization, matrix-free methods, Nyström and random-feature approximations, inducing points, and deep kernels all change the tradeoff. The model tested here makes a different move. It chooses a finite learned-center kernel expansion as the architecture, trains it by minibatches, and keeps its prototypes explicit. That does not inherit every fixed-kernel theorem, but it gives the risk score named parts that can be inspected.

This post builds the model and runs the trial in one place. The hidden layer is a bank of positive-definite kernel evaluations against learned prototype patients; a linear readout turns those scores into log-risk, and ordinary minibatched gradient descent trains the centers and readout. The benchmark then puts that architecture beside the models a survival statistician would reach for on five datasets. The question is not whether the white-box model wins every table. It is whether its explicit arithmetic remains useful at a measured predictive cost. This is a research illustration of interpretable survival modeling, not a clinical tool.

What is actually on trial#

What claim is worth five datasets and a battery of baselines? Not a leaderboard place. The thing under test is an existence proof, and it is worth stating plainly before any number arrives so that every number can be read against it.

You can build a network from positive-definite kernel evaluations against learned prototype patients and train it by ordinary minibatched gradient descent on the Cox partial likelihood. At a layer with a linear readout, the risk score has an exact finite arithmetic decomposition over its current centers. That does not put the full learned-center model in one fixed RKHS or grant it a classical fixed-kernel norm bound. Row edits and raw resemblance scores remain inspectable operations whose selectivity, calibration, and shift-detection value must be measured. This trial measures them.

Against that claim, the success criterion is not “beats Random Survival Forest.” It is “trains like a network, and ranks patients about as well as the classical machines that come with none of this structure.” Landing in the same interval as Cox and the forest, across a spread of real datasets, is the result, because it shows the trade is real: you can have the convenience of deep learning and the legibility of a kernel machine at the same time, and the accuracy you give up for it is small enough to see and to price. Where the white-box model costs a point or two, we say so and show where.

So the rest of the post is a trial with two halves. First, viability: can it hang with the baselines on concordance? Then, diagnostics: which proposed benefits of the explicit prototype structure survive direct measurement?

Build the score and expose its reasons#

For patient covariates $x$, event indicator $e_i$, and observed time $t_i$, DeepSurv trains one scalar log-risk $h(x)$ with the Cox partial loss

Only the trunk changes. The prototype version evaluates

Each $w_u$ is a synthetic patient profile in standardized covariate coordinates. The term $a_u\phi_u(x)$ is the exact additive contribution of that prototype to the current log-risk score. It is not a causal effect of a clinical variable, and the learned center may fall outside the data manifold; both issues are inspectable rather than hidden.

Deleting a row from the readout removes exactly that arithmetic term. That makes cohort edits reproducible, but it does not make them equivalent to retraining on a dataset with those people removed: the other centers and coefficients still encode the original optimization. The honest test is to perform the edit, measure selectivity, and state what remains fixed.

The lineup#

Who is the white-box model actually up against? Five datasets, chosen because they load without any private download and span very different survival regimes. METABRIC (breast cancer, n=1903, 42% censored) and GBSG (breast cancer recurrence, n=2232) are the mid-sized clinical staples. SUPPORT (critically ill hospitalized adults, n=8873) is larger and harder to rank. WHAS500 (heart-attack survival, n=500) is small and heavily censored, a stress test for anything data-hungry. FLCHAIN (a free-light-chain mortality cohort, n=6521, 70% censored) is nearly separable and where every model does well. Different sizes, different censoring, different signal.

Every model sees the same split, the same standardization, six seeds, and bootstrap confidence intervals on the held-out concordance index. The baselines are the ones a survival statistician reaches for: classical Cox proportional hazards, penalized Cox with an elastic-net path (Simon et al., 2011), and Random Survival Forest (Ishwaran et al., 2008), a strong nonlinear ensemble. Then two neural models at matched capacity: a standard ReLU DeepSurv (Katzman et al., 2018), the black box, and the Yat DeepSurv on trial.

One methodological point decides whether the comparison is fair, and it is worth naming because it is easy to get wrong. The two neural nets do not share a learning rate. A convergence probe found the Yat trunk and the ReLU MLP prefer different rates, so pinning them to one rate would be measuring an optimizer artifact, not a model. Each net instead sweeps its own learning rate over a small grid, picks the best by inner validation, and stops at its best epoch, all without ever touching the test set. The classical baselines have no learning rate, so they anchor the whole comparison. Every number below is from one run on Kaggle.

The viability verdict#

Here is the whole trial in one panel: for each dataset, every model’s held-out concordance index as a point, with its bootstrap 95% interval as a whisker. The shaded band is the classical baselines’ envelope, so “landing in the pack” is a literal thing you can see. Switch datasets and watch where the Yat point falls.

Read the panel and it says two different things on two groups of datasets. On METABRIC, SUPPORT, and GBSG the Yat model sits squarely inside the pack: 0.621 against a best-of-0.638 on METABRIC, 0.606 against 0.612 on SUPPORT, 0.662 against 0.676 on GBSG, whiskers overlapping every baseline. This is the existence proof landing: a kernel machine trained by gradient descent, ranking patients as well as classical Cox and within a whisker of a tuned forest, while carrying a theory none of them have.

On WHAS500 and FLCHAIN it costs something, and that cost is part of the result. WHAS500 is the sharpest: Cox reaches 0.765 and the Yat model 0.679, a real gap on a small, heavily censored cohort where the linear model’s strong prior pays off. On the near-separable FLCHAIN the field is bunched near 0.92 and the Yat model trails by about two points at 0.909. The five datasets locate the trade: competitive on three, clearly worse on two.

That is the entire viability claim: not a champion, but a member of the field in good standing that happens to bring a theory with it. The rest of the post is that theory.

How many prototypes, and does the seeding matter?#

A prototype model has a knob, the number of prototypes K, and a choice about where to place them at initialization. A sweep tests whether the result depends on one lucky setting.

The curve is nearly flat, which is the reassuring answer. Sweeping K from 6 to 48 moves the concordance index by only about a hundredth on either dataset; the model is not perched on a lucky value of K. K-means seeding and random seeding land close on average, with k-means the steadier choice (on METABRIC at K=6, k-means holds 0.629 against random’s 0.605 with roughly twice the seed spread). The reading is that a handful of well-placed prototypes already captures the signal, and adding more neither helps nor much hurts, which is the capacity story the kernel norm predicts. The knob is safe.

Diagnostic one: calibration#

A concordance index tells you the ranking is right; it says nothing about whether the numbers mean anything. An oncologist does not act on a rank, she acts on “70% chance of surviving five years,” and a model can rank perfectly while its probabilities are badly off. This is where the survival curve, not the risk score, is on trial, and it is the first thing the kernel structure lets you inspect directly.

On most datasets the two curves track closely: low-risk, mid-risk, and high-risk thirds separate, with reliability gaps of 0.022 on GBSG, 0.035 on SUPPORT, and 0.008 on FLCHAIN for the Yat model. METABRIC is looser at 0.078, and WHAS500 is the exception, with a gap of 0.265. These plots measure calibration; kernel structure does not guarantee it. Prototype terms may help investigate a failure, but the reliability curve is available for any probabilistic survival model.

Diagnostic two: does peak response detect a stranger?#

An uncalibrated model may hand out an extreme risk score far from its training distribution. This architecture exposes a candidate diagnostic: the largest raw activation over prototypes. For fixed centers the Yat response remains bounded as an input moves far away, although it is not a purely radial bump and its scale depends on norms, $b$, and $\varepsilon$. Whether kernel-max separates a relevant shift is therefore an empirical question, not a built-in defense.

The picture is mixed, and the mix is the interesting part. Shuffling each covariate independently, which breaks the joint distribution while preserving every marginal, is caught mildly (AUROC around 0.66 on METABRIC and GBSG): a real but subtle shift the meter half-notices. Synthetic far-tail patients, pushed several standard deviations out, are caught strongly on METABRIC (0.63) but, revealingly, not on GBSG or FLCHAIN, where pushing a normalized covariate far can still land near some prototype, so the meter reads them as familiar. And a real held-out subgroup, the top decile of the most variable covariate, often scores below 0.5, which reads at first like a failure and is actually a finding: that subpopulation is not strange to the kernel, it sits comfortably inside the learned field. The one true cross-dataset test that could run, scoring WHAS500 patients through the SUPPORT model, gives a kernel-max so low the separation is near-total, the strong signal you want when a genuinely foreign cohort walks in.

So the abstention property is real but not a magic OOD detector, and the previous post claimed too much by showing only the easy synthetic case. What it actually does: the kernel-max reliably flags a genuinely different data source and a strong synthetic push, mildly flags a broken joint distribution, and correctly does not flag an in-distribution subgroup that only looks extreme on one axis. That is a resemblance meter behaving like one, and it is a capability the Cox model and the forest simply do not have.

Diagnostic three: the prototypes against the clinic#

The whole legibility story rests on the prototypes being things a clinician could actually reason about. A prototype is a point in patient space, so in principle it is a synthetic patient, but nothing forces a learned point to be a plausible one; gradient descent could park a prototype in a math-useful but clinically absurd region. So the last question of the trial is a reality check: de-normalize every prototype to clinical units and ask whether it lands inside the range real patients occupy.

Eight of the twenty-four prototypes land fully inside the training range on every covariate, and the median prototype sits close to a real patient (distance 1.45 in nine-dimensional standardized space). The rest have one or two covariates that stray a little past the observed edge, usually a treatment indicator that a continuous prototype cannot represent cleanly, but their excursions are small (a fraction of a standard deviation) rather than wild. So a third of the prototypes are clinically plausible in the strict sense and most of the remainder are close, near-real profiles with a small excursion on a binary axis. That lands in a legible middle, and the useful part is that you can walk it prototype by prototype and say which is which. No forest offers that, because a forest has no prototypes to walk.

The scoreboard, and what it is not#

Here is the full concordance-index table, every number traced to the trial bundle, so the reader can audit the “in the pack” claim directly rather than take the forest plot on faith.

(Mean held-out concordance index over six seeds; higher is better, chance is 0.5. Full mean, standard deviation, AUC, integrated Brier, and bootstrap intervals are in the run bundle kgl_blog-deepsurv-trial-v2.)

What the trial settled#

Go back to the tension we opened on. A kernel machine gives you a theory and demands a solve that will not scale or compose. A deep network gives you the training and takes away the reasons. The claim of this whole series has been that the Yat kernel dissolves that choice, and this trial is the claim made to stand up in front of the baselines a skeptic would name.

The white-box survival model is viable on three of the five datasets and clearly worse on two. Its direct structural benefit is an explicit bank of prototype-patient scores whose arithmetic can be inspected. The calibration plots, row edits, and shift probes are experiments on top of that structure, with mixed outcomes and no clinical validation. The existence proof that survives is modest: a learned-center kernel survival network can be trained by minibatches and remain competitive in some regimes while exposing named internal references.

@misc{bouhsine2026survivalmodelontrial,
  author       = {Bouhsine, Taha},
  title        = {The White-Box Survival Model on Trial},
  year         = {2026},
  month        = {jul},
  howpublished = {\url{https://tahabouhsine.com/blog/survival-model-on-trial/}},
  note         = {Blog post, Records of the !mmortal Data Scientist}
}