Survival models predict how long until an event happens, such as death after a cancer diagnosis, when many patients are still alive at the end of follow-up. These "censored" patients have no observed event yet. Models of this kind are trained in one set of places and then deployed in others. The usual workhorse is the Cox proportional hazards model. It splits risk into a shared "baseline hazard," meaning how the event rate evolves over time for a reference patient, and a patient-specific multiplier.
Its training objective, the partial likelihood, cancels the baseline hazard entirely. So a Cox model can rank patients well in a new region while being badly wrong about absolute survival there. If a new state has different treatment access or follow-up practice, its whole survival curve shifts or changes shape, and nothing in the training signal notices. We propose to predict the baseline hazard of an unseen region from publicly available area-level data, such as census, environmental and built-environment descriptors, plus a smooth spatial residual. The result works like a stratified Cox model, which fits a separate baseline per region, except that it can produce a stratum for a region with no outcome labels at all.
The core claim is testable and falsifiable. Under leave-one-region-out evaluation on SEER (the US cancer registry program), we expect three things. First, ranking metrics such as the concordance index (C-index, the probability that the model orders two patients' event times correctly) will barely separate existing methods. Second, absolute-risk metrics such as the integrated Brier score and calibration of 5-year survival will separate them sharply. Third, almost all of that separation will come from modeling region-specific baseline hazard shape, not from richer patient-level risk functions.
What is already done, and what is not
Three nearby lines of work each solve one piece of this problem.
The first adds flexibility to the patient-level risk function. DeepSurv replaces the linear risk term with a neural network trained on the Cox partial likelihood. Neural Frailty Machine adds a random multiplicative "frailty" term and a fully neural variant that relaxes proportionality. Both assume one population drawn independently from a single distribution, with no notion of place.
The second transfers survival models across domains. Multi-Source Survival Domain Adaptation aligns features adversarially using a discordance-based distance between domains. Single-Domain Generalization for Multimodal Cross-Cancer Prognosis rebalances image and genomic features across cancer types. Both treat domains as discrete cancer types, and both are evaluated by C-index, which ignores the baseline hazard by construction.
The third handles geography. Deep Spatial Domain Generalization builds a spatial nearest-neighbor graph and uses a hypernetwork, a network that outputs the weights of another network, to produce location-specific task models at unseen coordinates. It reports an LA PM2.5 error of 9.40 against 12.44 for pooled training, but only for regression and classification. Federated Cox learning on SEER shows that heterogeneity across 10 states hurts plain federated averaging. It still treats states as private silos with no external data and no prediction for unseen states.
The obvious combination, a spatial hypernetwork that generates Cox risk weights, is not in the file. We nonetheless think it targets the wrong quantity. Any change to the relative-risk function leaves the baseline cancelled out, so it cannot repair absolute calibration in a new region. We keep it as a baseline rather than as the proposal.
Two results constrain the design. The Makassar dengue study found that district-level spatial frailty became non-significant once clinical covariates entered the model. The Nanjing asthma cohort linked multi-source exposure estimates into a time-dependent Cox model and got a PM2.5 hazard ratio of 1.01. Together they warn that area-level signal can be small once patient covariates are in. Our evaluation is therefore built so that a null result is visible and reportable, not hidden.
Core Claim
Geographic variation in survival has two parts.
The first part is proportional. It multiplies every patient's hazard in a region by the same factor at all follow-up times. A pooled neural Cox model can already learn this if area descriptors are concatenated to the inputs, because risk sets mix patients from several regions.
The second part is non-proportional. The baseline hazard's shape over follow-up time differs by region, for example through excess early mortality where treatment access is poor, or through different follow-up and censoring practice. No pooled partial-likelihood model can learn this part. A classical stratified Cox model handles it only for regions present in training.
We claim that, in multi-state cancer registries, the non-proportional part is large enough to matter for absolute risk. We further claim it can be predicted for unseen regions from external area-level data.
The model writes the hazard for patient covariates $x$ in region $s$ with area descriptors $d_s$ as
$$\log h(t \mid x, s) = \sum_{k=1}^{K} \mathbb{1}[t \in I_k]\,\big(\alpha_k + \gamma_k(d_s)\big) + g_\theta(x) + u(s),$$
with the terms defined as follows:
- $I_1,\dots,I_K$ are follow-up time intervals.
- $\alpha_k$ is a pooled log baseline hazard on interval $k$.
- $\gamma_k(\cdot)$ is a small network giving the region's interval-specific deviation from that baseline, which is what makes regions non-proportional.
- $g_\theta$ is a DeepSurv-style patient risk network shared across regions.
- $u(s)$ is a Gaussian-process residual over region centroid coordinates. It captures spatially smooth effects that the descriptors miss and can be predicted at a new location from its neighbors' fitted values.
Because the baseline is piecewise constant, the model is trained by full likelihood rather than partial likelihood. This is exactly what lets the baseline be learned and transferred. The descriptors $d_s$ come from integrating several data sources. We would draw on decennial census and American Community Survey (ACS) demographics as used in Transit for All, area deprivation indices like the Social Vulnerability Index used in the Miami street-view study, and gridded environmental exposures aggregated to the registry's reporting unit.
Research Program
The work has three stages.
Stage 1: measure the non-proportionality. On SEER cohorts, we test whether region effects violate proportional hazards. We fit a stratified Cox model by state, compare it with a model using state dummy variables, and apply standard residual-based proportionality tests to the region terms. If region effects turn out proportional, the core claim fails at this stage. Concatenating descriptors into DeepSurv would then be the right answer, and we would report that.
Stage 2: build and compare the transfer model. We fit the model above and compare it to the baselines below, using leave-one-state-out evaluation across the states in the SEER research release. The protocol follows the 10-state setup of the federated Cox study. Where county identifiers are available, we add a finer leave-one-county-cluster-out split.
Stage 3: semi-synthetic test with known ground truth. We keep real SEER covariates and real area descriptors but simulate event and censoring times. A tunable knob sets how much of the regional effect is proportional and how much changes baseline shape, and censoring rates drift with geography. This lets us verify that each component recovers what it should and nothing it should not.
Censoring differs across registries, so absolute-risk metrics on a held-out region use inverse-probability-of-censoring weights estimated from that region's own censoring distribution. Uncensored-looking errors therefore cannot reward a model that matches the training regions' follow-up pattern.
Baselines
The naive reference point is a pooled Kaplan–Meier curve, which ignores both covariates and place, together with a pooled linear Cox model using its Breslow baseline estimate (a nonparametric estimate of the baseline hazard). This is what a clinician applying a published nomogram in a new state effectively does. Every method must beat it on integrated Brier score in the held-out region.
The strongest existing competitor for ranking is a pooled DeepSurv given the same patient covariates. On a random split of SEER gastrointestinal stromal tumor patients, the SEER-based DeepSurv study reports a cancer-specific C-index of 0.825 and a Brier score of 0.142, against 0.761 for linear Cox and 0.789 for a random survival forest. We found no reported result on Lacuna for any survival model under leave-one-state-out SEER evaluation, so we will re-establish this baseline under our splits.
The strongest competitor for our specific question is the same pooled DeepSurv with the area descriptors $d_s$ and coordinates concatenated to its input. It has identical data access to our model and differs only in forcing every regional effect to be proportional. It gets the same hyperparameter search budget, the same early stopping on a held-out training region, and the same random seeds.
We also include four further comparisons:
- Spatial hypernetwork. A SIGNN-style hypernetwork adapted to output DeepSurv weights per location and trained on partial likelihood. This is the obvious idea at the intersection.
- Discrete domain adaptation. MSSDA with states as discrete source domains.
- Neural frailty. Neural Frailty Machine in its proportional-frailty form, pooled.
- Spatial frailty Cox. A Cox model with a Gaussian-process frailty over coordinates. This is the continuous-space analogue of the Leroux prior favored in the Makassar study.
The ablation that isolates the key ingredient removes $\gamma_k(d_s)$ and keeps everything else. The descriptors then enter only proportionally, and $u(s)$ becomes the sole regional term. A second ablation removes $u(s)$ to measure how much the external data carries on its own.
Beating the baselines means two things. Our model must match the strongest baseline's held-out C-index within a small margin. It must also reduce held-out integrated Brier score and 5-year calibration error relative to the concatenated DeepSurv, in a majority of held-out states, with confidence intervals obtained by bootstrapping over patients within each state.
Why not just…?
Why not just concatenate the area descriptors into DeepSurv? This is the real threat, and the proposal is designed around it. If regional differences in SEER are proportional, the concatenated model captures them through mixed-region risk sets and will tie with ours. Stage 1's proportionality tests and the $\gamma_k$ ablation settle the question directly. Evidence that naive concatenation can underperform exists, though not for this case. Transit for All reports learned fusion improving MAE by 11.74% over raw concatenation, and the cross-cancer study found concatenated multimodal models generalizing worse out of domain. Neither result concerns baseline hazards, so we treat this as an open empirical question.
Why not just fit a stratified Cox model or region dummies? Both give exact per-region baselines but have no value for a region absent from training. At most they could borrow the average stratum, which is the naive baseline. The metformin SEER-Medicare analysis is typical of this practice: it adjusts region away as a nuisance rather than predicting it. Our model reduces to a stratified Cox model when $d_s$ uniquely identifies training regions, so the comparison on seen regions also checks that we lose nothing there.
Why not just use a spatial hypernetwork or domain adaptation on the risk function? Both act on the relative-risk term, which the partial likelihood makes blind to the baseline. They can only improve ranking. MSSDA is evaluated purely on C-index for this reason. If our hypothesis is right, these methods will show small C-index differences from pooled DeepSurv and no calibration gain. If they instead close the calibration gap, the hypothesis is wrong.
Why not just use a spatial frailty model? Frailty is a single multiplier per location, so it is proportional by construction. The Makassar result shows that such terms can vanish after clinical adjustment. We include a Gaussian-process frailty baseline, and our $u(s)$ term contains it as a special case. Any gain over it must therefore come from time-varying baseline deviations.
Proposed Experiments
The decisive experiment is leave-one-state-out on SEER breast cancer, chosen to match the setting of the federated study. We will repeat it on a second, higher-mortality site such as lung cancer, where we expect regional treatment access to shape early hazards more strongly. Metrics are:
- Harrell's and time-dependent C-index.
- Integrated Brier score with held-out-region censoring weights.
- Calibration slope and absolute error of predicted versus held-out Kaplan–Meier survival at 1, 3 and 5 years.
The outcome pattern that would confirm the claim has three parts. Ranking metrics are close across all neural methods. Our model and the $\gamma_k$ ablation differ substantially on absolute-risk metrics. The size of that gap correlates across held-out states with the Stage 1 non-proportionality statistic.
The semi-synthetic experiment provides the second check. As the simulated non-proportional share goes to zero, our advantage over concatenated DeepSurv should vanish. Under geographically drifting censoring, the weighted evaluation should still rank methods correctly.
A last diagnostic regresses the fitted $\gamma_k(d_s)$ curves on individual descriptors. It asks whether area-level environmental exposures carry baseline-shape information once demographics are included, or whether they wash out as the Nanjing exposure signal did. We will report either answer.
References
- Deep Survival: A Deep Cox Proportional Hazards Network
- Neural Frailty Machine: Beyond proportional hazard assumption in neural survival regressions
- Multi-Source Survival Domain Adaptation
- Single-Domain Generalization for Multimodal Cross-Cancer Prognosis via Dirac Rebalancer and Distribution Entanglement
- Deep Spatial Domain Generalization
- Addressing Data Heterogeneity in Federated Learning of Cox Proportional Hazards Models
- Bayesian Spatial Survival Models for Hospitalisation of Dengue: A Case Study of Wahidin Hospital in Makassar, Indonesia
- Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China
- When aging meets digitisation: can urban digital technology innovations enhance the well-being of older adults?
- An Integrated Causal Inference Framework for Traffic Safety Modeling with Semantic Street-View Visual Features
- The development of a prediction model based on deep learning for prognosis prediction of gastrointestinal stromal tumor: a SEER-based study
- Transit for All: Mapping Equitable Bike2Subway Connection using Region Representation Learning
- Examining the relationship between metformin dose and cancer survival: A SEER-Medicare analysis