Fit a survival model to censored clinical outcomes

Hand Claude Code a tidy table of patient covariates plus a time-to-event column and an event/censoring indicator; get back a fitted Kaplan-Meier baseline, a Cox proportional-hazards model, a Random Survival Forest, honest cross-validated concordance, and the risk stratification that goes into a prognosis paper.

   
Problem class Data analysis
Subject areas Translational Medicine
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Time-to-event data is the backbone of clinical prognosis work — overall survival, progression-free survival, time-to-readmission. The analysis is standard but easy to get subtly wrong: censoring must be encoded as a structured (event, time) outcome (not a plain float), the concordance index must be evaluated out-of-sample (an in-sample c-index flatters every model), the proportional-hazards assumption needs checking before a Cox hazard ratio is trustworthy, and tree-based models like Random Survival Forests need the right censoring-aware splitting criterion. Every group rebuilds this boilerplate, and a misencoded outcome silently corrupts the whole analysis. Solved looks like: hand the agent a table, get a Kaplan-Meier curve, a Cox model with checked assumptions, an RSF for non-linear effects, and an honest held-out c-index — with every modeling choice written down.

  1. Install the scikit-survival skill in Claude Code:

    /plugin marketplace add K-Dense-AI/claude-scientific-skills
/plugin install scikit-survival@claude-scientific-skills

    The skill wraps the GPL-3.0 scikit-survival (sksurv) library. Confirm pip show scikit-survival returns a version ≥ 0.23.

  2. Put a tidy table in your project. One row per subject, numeric/categorical covariate columns, a duration column (follow-up time), and a binary event indicator (1 = event observed, 0 = censored). State the time unit explicitly — most survival bugs trace back to mixed units or to treating censored rows as events.

  3. Invoke the skill with the file path and the outcome encoding. A minimal prompt:

    Run the scikit-survival skill on data/cohort.csv. Outcome columns:
`os_months` (follow-up time, months) and `dead` (1=death, 0=censored).
Covariates: age, stage, ecog, treatment_arm, biomarker_high.
1. Build the structured survival outcome (Surv.from_arrays).
2. Plot the Kaplan-Meier curve overall and stratified by treatment_arm,
   with a log-rank p-value.
3. Fit a CoxPHSurvivalAnalysis model; report hazard ratios with 95% CIs.
4. Fit a RandomSurvivalForest (n_estimators=300, min_samples_leaf=15).
5. Report Harrell's c-index for both models via 5-fold cross-validation
   (concordance_index_censored on held-out folds, not in-sample).
Save the fitted-metrics table to results/survival_metrics.csv and the
KM + RSF figures to results/.

  4. Check the proportional-hazards assumption before trusting the Cox HRs. Ask the agent to inspect scaled-Schoenfeld-style residuals or stratify on any covariate that violates PH. A Cox HR is only meaningful if the hazards are proportional.

    Check the proportional-hazards assumption for each Cox covariate.
For any covariate that violates it, refit stratifying on that
covariate and show how the remaining HRs shift.

  5. Stratify into risk groups for the clinical readout. Use the model’s risk score to split the cohort (e.g., tertiles), then re-plot Kaplan-Meier per risk group with a log-rank test — this is the figure clinicians read.

  6. Hand off. The metrics CSV and figures drop into a manuscript. For external validation, run the saved model on a second cohort table in the same conversation and compare held-out c-index.

Why this assembly

Rung 2. scikit-survival is the validated Python survival-analysis stack (Cox, RSF, Gradient Boosted survival, c-index, integrated Brier score); the skill is a thin wrapper that pins the right idioms — structured Surv outcomes, censoring-aware scoring, out-of-sample evaluation. Plain Claude Code can write sksurv from memory but tends to drift on outcome encoding (the single most common survival bug) and on whether the c-index it reports is in-sample or held-out — exactly what the skill prevents. There is no need for a multi-tool harness or an autonomous system: this is one well-defined analytical task and a skill is the right grain.

Availability

Fully open. scikit-survival is GPL-3.0 and the K-Dense skill wrapper is OSS on the K-Dense marketplace. Any current Claude plan suffices. The cohort table stays local — no institutional license or data-residency gate from the tooling itself. Note that the data (e.g., MIMIC, a hospital registry) usually carries its own access and IRB constraints; those are the user’s to satisfy.

Compute requirements

Laptop. A cohort of a few thousand subjects with a dozen covariates fits Cox and a 300-tree RSF, with 5-fold CV, in seconds to a couple of minutes on 8 GB RAM. RSF memory and wall-clock scale with n_estimators × n_samples; tens of thousands of subjects still finish in minutes. No GPU needed.

Evidence

Proposed. No peer-reviewed paper documents this exact assembly (Claude Code + the K-Dense scikit-survival skill) running a survival analysis end-to-end. The component evidence is robust:

  • scikit-survival itself — Pölsterl, JMLR 21(212):1–6 (2020) is the canonical reference for the library, implementing Cox, Random Survival Forests, Gradient Boosted survival models, and censoring-aware metrics.
  • Random Survival Forests + concordance as the standard clinical method (analogous) — recent prognosis studies use exactly this pattern: Zhang et al., Transl. Cancer Res. (2026) compares nomograms and RSF for triple-negative breast-cancer survival on SEER; Liu et al., Medicine (2026) screens 101 survival algorithms (including RSF, LASSO-Cox, survival SVM) by mean c-index for colorectal-cancer prognosis. These validate the method, not the Claude assembly.

No published comparison of LLM-driven survival analysis against a hand-coded sksurv script is known. The skill adds consistent outcome encoding and out-of-sample scoring; it does not change the underlying statistical method.

Alternatives considered

  • Plain Claude Code, no skill. Works for users fluent in sksurv/lifelines. Reach for the skill when you want the outcome encoding and the held-out-c-index discipline pinned across runs and collaborators.
  • lifelines (Python) or the R survival / survminer stack. Domain-standard alternatives; lifelines is friendlier for pure Kaplan-Meier and parametric AFT models, R for publication-grade KM plots. Neither has a Claude-installable wrapper today; use them outside Claude Code when you need their specific idioms.
  • PyHealth deep-learning survival. For very large EHR cohorts where deep models add value, the PyHealth skill targets neural clinical-prediction tasks. That is a rung-2-to-3 escalation; reach for it only when classical Cox/RSF underperforms on a large, high-dimensional cohort.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.