Zero-Shot Survival Prediction from EHR Data — What Tabular Foundation Models Mean for Healthcare and Insurance

Hi, I’m Keito Inoshita from Affectosphere Group.

For anyone managing ICU bed capacity, the question “when can this patient move to a general ward?” is the first question of every morning.

Experienced physicians integrate test values, comorbidities, and vitals into an intuition-driven estimate. The case for machine learning is that it can formalize that intuition — but in practice, training a reliable survival model requires substantial patient data, and most hospitals working in specific clinical subpopulations don’t have it.

A paper published on arXiv in June 2026 (Minh-Khoi Pham et al., arXiv:2606.12006) directly addresses this barrier. The research adapts three tabular foundation models — TabPFN, TabICL, and TabDPT — to clinical survival analysis by attaching survival-aware prediction heads, without task-specific training on the target dataset. TabDPT achieved a C-index of 0.856, outperforming DeepSurv by 1.4%.

Today’s post unpacks what this means, and what it looks like to apply it in ICU management, life insurance underwriting, and oncology.

---

3 Points for Today

1. Tabular foundation models — a fast-moving research direction in AI — have now demonstrated competitive survival prediction from clinical data without additional training.
2. Survival analysis is a practical framework for estimating “when an event happens,” not just “whether it happens” — and handling incomplete observations correctly.
3. Three concrete deployment scenarios with department-level and KPI-level specifics: ICU, life insurance, and chronic disease prognosis.

---

① What Tabular Foundation Models Are

For the past several years, foundation models have dominated image and text AI. Their application to tabular data — spreadsheets, databases, electronic health records — is newer and still actively developing.

TabPFN, TabICL, and TabDPT are representative models in this space, each pre-trained at scale on diverse tabular tasks. Their defining characteristic: they can generalize to new tabular prediction tasks without task-specific training, or with minimal fine-tuning.

The research question Pham et al. pursued was: can these models be adapted for survival analysis?

Survival analysis estimates “time until an event” — death, discharge, disease recurrence. The critical technical challenge is handling censored data: patients for whom the event has not yet been observed by the end of the study period. Getting censoring right is what distinguishes survival models from standard classification or regression.

The team attached survival-aware prediction heads to each foundation model and adapted them with censoring-aware loss functions — partial likelihood and ranking-based losses. This is the “survival-aware adaptation” of the paper’s title.

---

② Why TabDPT Outperformed DeepSurv

The evaluation metric is the C-index (Concordance Index): the proportion of all comparable patient pairs where the model’s predicted risk ranking matches the actual event order. 0.5 is random, 1.0 is perfect.

TabDPT achieved C-index 0.856 against DeepSurv’s 0.842 — a 1.4 percentage point improvement.

That gap may look small. In clinical terms, a model that more consistently ranks patients by actual risk trajectory has direct implications for which patients receive priority intervention. At 0.856 C-index, the ordering is correct in roughly 6 out of 7 comparable patient pairs.

The more striking detail is the zero-shot condition. DeepSurv and similar approaches train on the target dataset. The tabular foundation models in this study arrived pre-trained and were evaluated without additional task-specific training. The implication: facilities with limited labeled patient data may still achieve competitive performance.

---

③ Three Deployment Scenarios

Scenario 1: ICU discharge timing (ICU administrators)

The core operational problem in ICU management is estimating when a patient can safely transfer to a general ward. Too long increases cost and resource pressure; too early increases readmission risk.

Integrating a survival model trained on vitals, labs, and clinical history can produce patient-level estimates of “time until ICU-exit event” as survival curves — giving clinical teams a probabilistic discharge window rather than a point estimate.

Department mapping: hospital medical information systems team or ICU clinical informatics team, in coordination with intensivists. KPI: reduction in unplanned ICU length-of-stay extension rate; readmission rate within 48 hours of transfer.

Scenario 2: Life insurance underwriting (actuaries)

Life insurance underwriting assigns risk categories and premiums based on mortality likelihood. Current approaches are largely mortality-table-driven. Where EHR data is accessible for applicants, individual-level survival modeling opens the possibility of more granular risk stratification.

A survival model estimates the probability that a mortality event occurs within a specified time horizon at the individual level. The value of tabular foundation models here is that underwriting teams don’t need to build large training datasets for each risk subpopulation from scratch.

Department mapping: actuarial department or underwriting division. KPI: improvement in loss ratio; reduction in adverse selection rate; underwriting turnaround time.

Scenario 3: Cancer and chronic disease prognosis (hospital leadership and oncologists)

Cancer prognosis and chronic disease trajectory prediction — heart failure, CKD, COPD — are classical survival analysis applications. The friction for most hospitals has been the AI development cost: collecting clean data, training a model, maintaining it.

Tabular foundation models reduce the entry cost. If structured EHR data can be used as-is without extensive feature engineering, the data infrastructure requirement drops significantly. This makes the approach more accessible to mid-sized hospitals that lack dedicated AI teams.

Department mapping: oncology or cardiology clinical teams in partnership with the medical informatics department. KPI: C-index tracked quarterly on held-out patient cohorts; utilization rate of the model’s priority score in treatment planning rounds.

---

What to Keep in Mind Before Deploying

The C-index of 0.856 is a benchmark result on specific datasets. Clinical EHR data in practice varies substantially — different missingness patterns, variable definitions, and coding conventions. Assuming benchmark performance will transfer directly is premature.

“Zero-shot” means no additional training, which also means no learning from site-specific clinical patterns. For facilities with distinctive patient populations, combining the foundation model with minimal fine-tuning on local data is likely to improve calibration.

Interpretability is the other consideration. Survival model outputs in clinical workflows need to be explainable to the clinical team using them. A score without a clear interpretation of the driving features will struggle to gain trust in practice.

---

The Shift That This Paper Represents

The deeper story here is the same one that played out in image and text AI several years ago: from “train a model from scratch for every task” to “adapt a pre-trained foundation model to each task.”

That shift is now reaching structured clinical data. For healthcare organizations that hold EHR data, this means the barrier to deploying survival prediction models is lower than it was. Not zero — but meaningfully lower than training from scratch in a data-scarce clinical setting.

That’s the practical bet worth making as this space matures.

That’s all for today!

---

References

1. Minh-Khoi Pham et al. (2026). Tabular Foundation Models for Clinical Survival Analysis via Survival-Aware Adaptation. arXiv:2606.12006.

* This article was written in part with AI assistance and may contain inaccuracies.