OncologyConsultSurvivalLlm Agent

<!--
# COPYRIGHT NOTICE
# This file is part of the "Universal AI Agentic Skills" project.
# Copyright (c) 2026 MD BABU MIA, PhD <md.babu.mia@mssm.edu>
# All Rights Reserved.
#
# This code is proprietary and confidential.
# Unauthorized copying of this file, via any medium is strictly prohibited.
#
# Provenance: Authenticated by MD BABU MIA
-->



<!-- AUTHOR_SIGNATURE: 9a7f3c2e-MD-BABU-MIA-2026-MSSM-SECURE -->

---
name: 'oncology-consult-survival-llm'
description: 'Guide zero-shot or fine-tuned LLM workflows for predicting cancer survival from initial oncology consultation documents with leakage control and cautious reporting.'
measurable_outcome: 'Execute skill workflow successfully with valid output within 15 minutes.'
allowed-tools:
  - read_file
  - run_shell_command
  - web_fetch
---

# Oncology Consult Survival LLM

## Overview

This skill guides clinically cautious workflows for using general large language models to predict cancer survival from initial oncology consultation documents. It emphasizes survival endpoint definition, zero-shot versus fine-tuned evaluation, data leakage controls, calibration, uncertainty, and reporting that is appropriate for research or decision support rather than standalone clinical judgment.

Use this skill when the task involves oncology-document prognostic modeling and the user needs a defensible workflow grounded in the 2026 PubMed-indexed finding on fine-tuning versus zero-shot LLM prediction of cancer survival from initial consultation notes.

## When to Use This Skill

- Predicting survival, mortality, or time-to-event outcomes from initial oncology consultation documents.
- Comparing zero-shot prompting with fine-tuned LLM approaches for oncology prognosis tasks.
- Designing training, validation, or test splits for clinical-note survival prediction.
- Reviewing oncology prognostic modeling plans for leakage, endpoint ambiguity, label quality, or calibration gaps.
- Producing clinically cautious summaries of LLM-derived survival risk predictions.
- Auditing whether consultation-note text contains future information that would invalidate a baseline prediction task.

## Core Capabilities

1. **Survival endpoint framing**: Define the target outcome before modeling, including prediction time zero, follow-up window, event type, censoring rule, and whether the endpoint is binary survival status or time-to-event risk.

2. **Document cohort preparation**: Confirm that each input is an initial oncology consultation document and preserve metadata needed for temporal ordering, cancer type, disease stage, treatment intent, and follow-up ascertainment.

3. **Leakage control**: Exclude future information such as post-consultation treatments, later imaging, later pathology amendments, recurrence notes, death documentation, hospice enrollment after baseline, or follow-up outcomes embedded in copied text.

4. **Zero-shot workflow design**: Build explicit prompts that ask for structured prognostic outputs, cite only baseline document evidence, separate unknown from absent information, and avoid forcing a risk estimate when the note is insufficient.

5. **Fine-tuning workflow design**: Prepare de-identified training examples with stable labels, patient-level splits, development and held-out test sets, locked preprocessing, and no overlap across documents from the same patient.

6. **Evaluation and calibration**: Report discrimination, calibration, threshold behavior, uncertainty, and clinically relevant subgroup checks without inventing or overstating performance. Prefer calibration plots, Brier score, concordance-style summaries, or sensitivity and specificity when appropriate to the endpoint.

7. **Clinical caution and reporting**: State that outputs are model-derived prognostic estimates requiring clinician review and local validation. Avoid treatment recommendations unless the user separately requests guideline-grounded clinical decision support.

8. **Comparison interpretation**: When comparing zero-shot and fine-tuned models, separate performance from operational tradeoffs such as data requirements, reproducibility, site-specific adaptation, privacy burden, and robustness to documentation style.

9. **Comparative validation framework**: Compare fine-tuned and zero-shot models against simpler baselines using patient-level temporal splits, strict leakage controls, censoring-aware endpoints, and calibration assessment; seek external validation when available, and explicitly report when fine-tuning does not outperform zero-shot or simpler approaches.

10. **Zero-shot versus fine-tuning decision framework**: Select between zero-shot and fine-tuned approaches only after evaluating both with patient-level temporal splits, document leakage controls, censoring-aware endpoints, calibration, external validation where available, and clinically relevant subgroup analysis; report comparative results cautiously, including uncertainty, limitations, and whether evidence supports use beyond research or clinician-reviewed decision support.

11. **Direct adaptation comparison framework**: Compare zero-shot prompting and task-specific fine-tuning on the same endpoint, eligible cohort, initial-consultation document cutoff, preprocessing, and held-out patients. Enforce the cutoff before text extraction and audit copied-forward or dated content for post-baseline leakage; document class imbalance and use imbalance-aware evaluation rather than accuracy alone. Preserve temporal validation and, when available, external-site validation; compare calibration as well as discrimination, report confidence intervals or other uncertainty estimates, and disclose unstable subgroup results. Prefer prompting when it provides sufficiently calibrated, reproducible performance under these validations with lower data and governance burden; choose task-specific adaptation only when it yields a reliable, clinically meaningful improvement that persists across temporal and external evaluation.

12. **Evidence-informed model selection**: Use the 2026 PubMed-indexed fine-tuning versus zero-shot survival prediction study as evidence to require direct comparison of candidate general LLM approaches on initial oncology consultation documents, with post-baseline leakage prevention, held-out patient validation, calibration review, and cautious clinician-reviewed reporting before any research or decision-support use.

13. **Fine-tuning versus zero-shot evaluation module**: Evaluate fine-tuned and zero-shot general LLMs on the same initial-consultation survival prediction task with locked train/test cohort separation, patient-level non-overlap, post-baseline leakage controls, calibration assessment, censoring-aware reporting, and conservative language that limits outputs to research or clinician-reviewed decision support unless locally validated.

14. **Initial-consultation comparison safeguards**: When comparing fine-tuning versus zero-shot survival prediction from initial oncology consultation documents, use the same baseline document cutoff and endpoint definition for both approaches; prevent leakage from future notes, outcomes, treatments, or copied-forward post-baseline content; split cohorts by patient and, when feasible, by time and site; include calibration review and censoring-aware reporting; and state that any clinical use requires clinician review and local validation.

15. **Finding-specific comparison checklist**: For survival prediction from initial oncology consultation documents, compare zero-shot and fine-tuned LLMs only after locking the endpoint definition and prediction time zero, applying identical baseline document cutoffs, preventing post-consultation leakage, evaluating calibration and uncertainty on held-out patients, seeking external validation before generalizing beyond the source setting, and reporting outputs as cautious clinician-reviewed research or decision-support estimates rather than definitive survival predictions.

16. **Study-specific non-deployment guardrails**: When applying the 2026 fine-tuning versus zero-shot oncology consultation survival prediction finding, preprocess clinical documents to remove or flag copied-forward, dated, outcome-revealing, or post-consultation text before comparison; lock the survival endpoint and prediction time zero; review calibration and subgroup performance on held-out patients; and frame results as prognostic research signals requiring local validation rather than deployment-ready clinical predictions.

17. **Model-selection and validation pattern**: Treat the 2026 fine-tuning versus zero-shot oncology consultation survival study as a pattern for selecting between general LLM approaches only after using the same initial-consultation baseline, enforcing patient-level separation and leakage controls, applying censoring-aware survival evaluation, checking calibration and clinically relevant subgroups, and reporting results in conservative language that avoids definitive survival claims.

18. **PubMed 42004490 comparison update**: For initial oncology consultation-note survival prediction, treat fine-tuning versus zero-shot general LLM use as a head-to-head evaluation question rather than assuming one approach is superior. Lock the survival endpoint and prediction time zero before modeling, prevent future-outcome leakage, assess calibration on held-out patients, and limit outputs to cautious clinician-reviewed research or decision-support estimates.

19. **Fine-tuning versus zero-shot survival guidance**: When using general LLMs to predict cancer survival from initial oncology consultation notes, require parallel zero-shot and fine-tuned evaluation on the same endpoint and patient-level cohorts; perform leakage checks for post-consultation outcomes, treatments, copied-forward content, and dated future evidence; stratify cohorts by clinically relevant factors such as cancer type, stage, and treatment intent when available; assess calibration and report uncertainty rather than point estimates alone; and state that outputs are not standalone clinical predictions without clinician review and local validation.

20. **Consultation-note survival study controls**: For work informed by the 2026 fine-tuning versus zero-shot study, prespecify the survival endpoint, censoring rule, and prediction time zero before text extraction; use patient-level cohort splits with no document overlap; compare zero-shot and fine-tuned LLMs against a clearly defined baseline comparator on the same eligible cohort; audit post-consultation leakage; evaluate calibration and uncertainty; and report findings cautiously as initial-consultation-note research signals rather than standalone survival predictions.

21. **PubMed finding integration**: Incorporate the 2026 ESMO Real World Data Digit Oncol finding as support for comparing zero-shot and fine-tuned general LLMs on initial oncology consultation documents without assuming either approach is clinically sufficient. Define the survival endpoint and prediction time zero before modeling, enforce post-consultation leakage checks, assess calibration on held-out patients, and report outputs with cautious clinician-reviewed language requiring local validation.

22. **Tumor-board survival comparison guidance**: Choose zero-shot versus fine-tuned LLMs through a prespecified head-to-head validation on the same initial oncology consultation-note cohort rather than assuming either approach is superior. Freeze prediction time zero at the initial consult, remove or flag post-consultation outcomes, treatments, copied-forward dated evidence, and other future information, calibrate time-to-event outputs against held-out follow-up and censoring data, and present uncertainty explicitly for tumor-board review.

23. **Initial-consult survival study guidance**: Apply the PubMed 42004490 finding by splitting cohorts at the patient level with no train/test overlap, enforcing post-consultation leakage controls before text extraction, comparing fine-tuned and zero-shot LLMs against a defined baseline, using survival-specific and censoring-aware metrics with calibration review, and reporting prognostic outputs from initial consult documents cautiously as research or clinician-reviewed decision-support estimates.

24. **Chronology, subgroup, and governance gate**: For fine-tuning versus zero-shot survival prediction from initial oncology consultation documents, require patient-level train/test separation that respects document and outcome chronology when timestamps are available; audit post-baseline leakage, copied-forward future evidence, and label-derived text before modeling; assess calibration and clinically relevant subgroup errors on held-out patients; and document model governance, local validation needs, and cautious clinician-reviewed framing before any prognostic use.

25. **Zero-shot versus fine-tuned consult-note comparison**: Compare zero-shot and fine-tuned general LLM survival predictions from initial oncology consultation documents on prespecified, identical endpoints and patient-level cohort splits; control leakage by excluding post-consultation outcomes, later treatments, copied-forward future evidence, and label-derived text; evaluate calibration and report uncertainty on held-out patients; and interpret results cautiously as research or clinician-reviewed decision-support signals rather than standalone clinical predictions.

26. **Fine-tuning decision gate with drift checks**: Decide between zero-shot and fine-tuned survival prediction only after head-to-head validation on initial consultation notes with the same endpoint, prediction time zero, patient-level splits, and leakage controls. Check for cohort drift across training, validation, and intended-use cohorts, including cancer mix, stage, treatment intent, note source, site, and time period; reassess calibration under drift and report predictions cautiously as clinician-reviewed research or decision-support estimates requiring local validation.

27. **Consultation-document comparison protocol**: Compare fine-tuned and zero-shot general LLM survival predictions from initial oncology consultation documents under a shared protocol: lock prediction time zero, survival endpoint, censoring rule, and follow-up window; use patient-level validation splits with no train, validation, or test overlap and temporal or external splits when available; audit post-consultation treatments, outcomes, copied-forward future information, and label-derived text for leakage; check calibration and uncertainty on held-out patients; and report results cautiously as clinician-reviewed research or decision-support estimates requiring local validation rather than definitive survival predictions.

28. **Explicit zero-shot versus fine-tuned comparison guidance**: For general LLM survival prediction from initial oncology consult notes, evaluate zero-shot and fine-tuned approaches head-to-head on the same cohort, document cutoff, survival endpoint, censoring rule, and time horizon; control leakage from post-consultation outcomes, later treatments, future dated text, copied-forward content, and label-derived signals; assess calibration and uncertainty on held-out patients; seek external validation before applying results beyond the source setting; and report clinician-facing outputs cautiously as research or decision-support estimates requiring clinician review, not standalone survival predictions.

29. **Consultation-note cohort and prognostic-output safeguards**: For workflows informed by the 2026 fine-tuning versus zero-shot survival prediction study, construct the cohort from initial oncology consultation documents with patient-level identifiers, consultation dates, endpoint definitions, follow-up and censoring data, and no train/validation/test patient overlap; prevent leakage from post-consultation notes, outcomes, treatments, copied-forward dated text, and label-derived signals before either prompting or fine-tuning; compare zero-shot and fine-tuned models on the same eligible cohort and endpoint; review calibration and uncertainty on held-out patients; and report prognostic LLM outputs cautiously as clinician-reviewed research or decision-support estimates requiring local validation.

30. **Fine-tuning versus zero-shot prognostic governance**: Use the PubMed 42004490 finding to require a prespecified comparison of fine-tuned and zero-shot general LLM survival prediction from initial oncology consultation notes without assuming superiority; enforce leakage controls for future outcomes, post-consultation treatments, copied-forward dated text, and label-derived signals; check cohort drift across training, validation, test, and intended-use populations; use censoring-aware evaluation and calibration review; and communicate prognostic outputs in guarded language as locally validated, clinician-reviewed research or decision-support estimates rather than definitive survival predictions.

31. **Document-window constrained survival comparison**: For fine-tuning versus zero-shot survival prediction from initial oncology consultation documents, restrict model inputs to the prespecified initial-consultation window, define outcome censoring and follow-up before evaluation, audit leakage from future outcomes or post-consultation content, assess calibration on held-out patients, and report results cautiously as clinician-reviewed research or decision-support estimates rather than standalone clinical predictions.

32. **Fine-tuning versus zero-shot consult-study protocol**: Use the PubMed 42004490 oncology consultation survival study to guide leakage-controlled comparison of fine-tuned and zero-shot general LLMs only after defining the survival endpoint, prediction time zero, censoring rule, and follow-up window; preprocess documents to restrict inputs to initial-consultation content and flag copied-forward, dated, outcome-revealing, or post-consultation text; evaluate calibration and uncertainty on held-out patients; and report predicted survival risk cautiously as a clinician-reviewed research or decision-support estimate requiring local validation.

33. **Worked consultation-note comparison for tumor-board review**: For a worked fine-tuning versus zero-shot comparison on initial oncology consultation notes, define the survival endpoint, prediction time zero, censoring rule, and follow-up window before modeling; split cohorts at the patient level with no train, validation, or test document overlap and use temporal or site splits when available; restrict inputs to the prespecified initial-consultation window and audit future outcomes, later treatments, copied-forward dated text, and label-derived signals for leakage; review calibration and uncertainty on held-out patients; and present results cautiously for tumor-board review as clinician-reviewed research or decision-support estimates requiring local validation.

34. **Fine-tuning versus prompting evidence gate**: Use the PubMed 42004490 study as evidence that choosing fine-tuning versus zero-shot prompting for general LLM survival prediction from initial oncology consultation documents should be a prespecified comparison, not an assumption. Define the endpoint, prediction time zero, censoring rule, follow-up window, and patient-level cohort splits before modeling; enforce leakage controls for post-consultation outcomes or copied-forward future text; assess calibration on held-out patients; and report clinical-facing outputs cautiously as research or clinician-reviewed decision-support estimates requiring local validation.

35. **Fine-tuning versus zero-shot decision framework**: For cancer survival prediction from initial oncology consultation documents, compare zero-shot prompting and fine-tuning with a prespecified protocol that extracts only baseline clinical covariates available in the initial consult, locks patient-level train, validation, and test cohort splits before modeling, controls leakage from future outcomes, treatments, copied-forward text, or label-derived signals, evaluates calibration and prognostic uncertainty on held-out patients, and reports risk estimates cautiously as clinician-reviewed research or decision-support outputs rather than definitive survival predictions.

36. **Initial-consult train/test safeguard**: For fine-tuning versus zero-shot survival prediction from initial oncology consultation documents, use locked patient-level train/test document splits with no overlap, restrict inputs to baseline consultation text, audit future outcome, treatment, copied-forward, dated, and label-derived leakage before prompting or fine-tuning, evaluate calibration and clinically relevant subgroup performance on held-out patients, and report prognostic outputs cautiously as clinician-reviewed research or decision-support estimates requiring local validation.

## Inputs / Outputs

**Inputs**

- Initial oncology consultation documents or extracted consultation-note text.
- Cohort metadata such as patient identifier, consultation date, cancer diagnosis, stage, treatment intent, and source system.
- Survival labels or follow-up data with event date, censoring date, last known alive date, and endpoint definition.
- Modeling mode: zero-shot prompting, fine-tuning, or comparative evaluation.
- Constraints for de-identification, governance, protected health information handling, and permitted tools.

**Outputs**

- A structured workflow for survival prediction from baseline oncology consultation text.
- A leakage audit checklist tailored to the available documents and labels.
- Prompt templates or fine-tuning data schema suitable for the specified endpoint.
- Evaluation plan covering discrimination, calibration, uncertainty, and subgroup review.
- Clinically cautious reporting language that describes model outputs as research or decision-support signals, not definitive predictions.

## References

- Phaterpekar T, Zeng Z, Mali Y, Leung B, Ho C. "Investigating fine-tuning versus zero-shot learning for general large language models when predicting cancer survival from initial oncology consultation documents." ESMO Real World Data Digit Oncol. 2026 Jun. PubMed: https://pubmed.ncbi.nlm.nih.gov/42004490/
- https://pubmed.ncbi.nlm.nih.gov/42004490/