clinical-ai-modeling

name	clinical-ai-modeling
description	Plan and review clinical AI modeling workflows with EHR cohorts, PyHealth-style tasks, validation, bias checks, calibration, safety boundaries, and no-PHI handling. Use when users ask about clinical prediction models, risk scores, treatment-response ML, medical reports, or clinical model evaluation.
user-invocable	true
license	Apache-2.0

Clinical AI Modeling

Source adaptation: BioRouter-original synthesis inspired by Scientific Agent Skills clinical AI patterns, rewritten under Apache-2.0.

When to Use

Use this skill for clinical prediction, risk modeling, treatment-response modeling, EHR ML, report-generation evaluation, and clinical AI validation. Use clinical-decision-support-review when the output is a care recommendation or CDS text; use this skill when the core work is model design/evaluation.

Intake

Collect:

• Clinical task: diagnosis, prognosis, readmission, mortality, treatment response, phenotyping, summarization, or triage.
• Cohort definition, index time, prediction time, horizon, and outcome definition.
• Data source: OMOP, FHIR, CDW, claims, notes, imaging, labs, vitals, medications, or multimodal data.
• De-identification/PHI constraints.
• Model type and evaluation target.
• Deployment status: research prototype, retrospective validation, prospective validation, or operational CDS.

Cohort and Label Gates

1. Define index time before features.
2. Prevent label leakage from future notes, orders, procedures, outcomes, or discharge artifacts.
3. Capture inclusion/exclusion criteria and censoring.
4. Preserve code sets and vocabulary versions.
5. Separate development, validation, and external test cohorts.

Evaluation Gates

• Discrimination: AUROC/AUPRC where appropriate.
• Calibration: reliability curve, calibration slope/intercept, Brier score.
• Clinical utility: decision curves or threshold-specific confusion matrices.
• Fairness/subgroups: site, race/ethnicity, sex/gender, age, insurance, language, and disease severity where available.
• Robustness: temporal validation, site validation, missingness stress tests.
• Explainability: use only as supporting evidence, not proof of safety.

BioRouter Routes

• Use CDWAgent, UCSFOMOPAgent, or other BRXT extensions for authenticated data access.
• Use causal-identification-gates when causal treatment language appears.
• Use statistical-modeling-bayes for uncertainty and hierarchical validation.
• Use claim-evidence-integrity for manuscripts or model cards.
• Use regulatory-quality-systems for SaMD, design controls, or audit readiness.

Safety Boundaries

Never present an unvalidated model as clinical advice. Mark research-only outputs clearly unless prospective validation, governance, monitoring, and deployment controls are documented.

Output Shape

Return: task framing, cohort/label definition, model plan, leakage checks, evaluation table, subgroup plan, deployment/safety status, and open validation gaps.