# Censored isotonic teacher for neural survival heads

- ID: 1737
- Canonical URL: https://synthcore.org/idea/1737/censored-isotonic-teacher-for-neural-survival-heads
- API JSON: https://synthcore.org/api/idea/1737.json
- API Markdown: https://synthcore.org/api/idea/1737.md
- Verification status: unverified
- Source: [arXiv:2608.02914](https://arxiv.org/abs/2608.02914)
- Category: regularization
- Solves: accuracy, stability, generalization
- ML areas: mlp, loss, regularization, fine-tuning
- Math tags: statistics, optimization, probability, order-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Use Survival-IDR as a nonparametric calibration teacher for a neural conditional survival model when a covariate, risk score, or one-dimensional learned index has a known monotone relationship with event-time distributions. The teacher corrects the biased behavior of naive pooled Kaplan-Meier estimates under censoring and supplies distributional targets that are monotone across the ordered axis and coherent across every partition scale. Fine-tune the neural head against these targets while retaining the censored likelihood, testing whether calibration and monotonicity improve without sacrificing discrimination.

## Mathematical statement

For an ordered sequence of covariate groups indexed by 1 through m, let I=[r:s] denote a contiguous interval and let \widehat{\mathbb F}_{I}(y) be its Kaplan-Meier estimate of the conditional CDF at event-time threshold y. If d_{I,j} events occur at distinct observed time t_{I,j} and Y_{I,j} subjects are at risk immediately before that time, then \widehat{\mathbb F}_{I}(y)=1-\widehat S_I(y), where \widehat S_I(y)=\prod_{t_{I,j}\le y}(1-d_{I,j}/Y_{I,j}). Let \mathcal P_{r:s} be the set of all interval partitions of [r:s] into at least two nonempty contiguous intervals. S-IDR constructs \widetilde{\mathbb F}_{r:s}(y) inside the Cauchy mean value interval with lower endpoint L_{r:s}(y)=\max_{P\in\mathcal P_{r:s}}\min_{I\in P}\widetilde{\mathbb F}_{I}(y) and upper endpoint U_{r:s}(y)=\min_{P\in\mathcal P_{r:s}}\max_{I\in P}\widetilde{\mathbb F}_{I}(y). The paper proves L_{r:s}(y)\le U_{r:s}(y) by induction, so the recursive clamp \widetilde{\mathbb F}_{r:s}(y)=\min\{U_{r:s}(y),\max\{L_{r:s}(y),\widehat{\mathbb F}_{r:s}(y)\}\} is always defined. The neural adaptation uses T_i(y)=\widetilde{\mathbb F}_{B(i)}(y), where B(i) is the ordered score bin containing sample i, as a calibration target for the network CDF F_    heta(x_i,y).

## Key formulas

- $$\widehat{\mathbb F}_{I}(y)=1-\widehat S_I(y),\qquad \widehat S_I(y)=\prod_{t_{I,j}\le y}\left(1-\frac{d_{I,j}}{Y_{I,j}}\right)$$
- $$\widetilde{\mathbb F}_{r:s}(y)\in\bigcap_{P\in\mathcal P_{r:s}}\left[\min_{I\in P}\widetilde{\mathbb F}_{I}(y),\,\max_{I\in P}\widetilde{\mathbb F}_{I}(y)\right]$$
- $$L_{r:s}(y)=\max_{P\in\mathcal P_{r:s}}\min_{I\in P}\widetilde{\mathbb F}_{I}(y),\qquad U_{r:s}(y)=\min_{P\in\mathcal P_{r:s}}\max_{I\in P}\widetilde{\mathbb F}_{I}(y)$$
- $$\widetilde{\mathbb F}_{r:s}(y)=\min\left\{U_{r:s}(y),\max\left\{L_{r:s}(y),\widehat{\mathbb F}_{r:s}(y)\right\}\right\},\qquad L_{r:s}(y)\le U_{r:s}(y)$$

## Implementation notes

1. Integration point: use a discrete-time neural survival model whose logits produce F_theta(x_i,y_k) in [0,1] at K increasing event-time thresholds, or convert predicted hazards into a CDF. Choose an ordered scalar z_i, such as a clinically known monotone covariate, the model's scalar risk score, or a separately trained one-dimensional index. Sort calibration examples by z_i and divide them into bins with minimum size 30. Compute the S-IDR teacher only on a held-out calibration split so that targets are not obtained from the same examples used for fitting. 2. Pseudocode: for each threshold y_k, sort calibration samples by z; for every contiguous interval I=[r:s], compute risk counts Y_{I,j} and event counts d_{I,j}, then calculate the Kaplan-Meier estimate using the first formula. Process intervals from length 1 upward. Compute L=max over partitions P of min over intervals I in P of the already computed tilde-F_I value, and U=min over partitions P of max over intervals I in P of that value. Assign tilde-F_{r:s}=clip(hat-F_{r:s},L,U). A dynamic program over contiguous partitions or the paper's PAVA implementation can compute these bounds. Map the result to every sample in its bin to obtain T_i(y_k). Fine-tune with L=L_censored(theta)+lambda times the sum over i,k of w_{ik} BCE(F_theta(x_i,y_k),T_i(y_k)), where w_{ik}=0 for thresholds having fewer than 10 at-risk subjects. 3. Computed from the paper's mathematics: Kaplan-Meier interval estimates, all-partition lower and upper bounds, and the guaranteed nonempty clamp interval. Estimated empirically: the risk-score ordering, bin count, lambda, and tail weights. 4. First experiment: generate 20,000 samples with z uniformly distributed on [-1,1], event time Y following a Weibull distribution with shape 1.5 and scale exp(-z), and independent censoring tuned to 50 percent. Train a small MLP discrete-time hazard model with ordinary censored likelihood, with and without S-IDR distillation. Compare against a same-capacity DeepHit-style baseline. Measure integrated Brier score, expected calibration error at each threshold, negative log-likelihood, concordance, and monotonicity violations of F_theta across ordered z values. Success means lower Brier score and calibration error, fewer violations, and no meaningful concordance loss at equal training FLOPs. Repeat on SUPPORT or METABRIC with five random seeds.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.