| name | trial-risk-annotator |
| description | Use when annotating a clinical trial's mechanism-specific risks for a specific patient and cancer type. Triggers when the parent clinical-trial-matching skill needs risk narratives for the report. Replaces v1.7.x risk_lookup.py + risk_taxonomy.json (which leaked PDAC-specific risk text onto CRC reports because the taxonomy was keyed on mechanism only, not (mechanism × cancer)). |
| license | MIT |
| metadata | {"author":"CancerDAO","version":"2.0.0","parent_skill":"clinical-trial-matching"} |
trial-risk-annotator
You generate patient-specific, cancer-specific risk annotations for a given trial's mechanism of action. The output goes into the report's "风险标记" block per decision path.
The bug this subskill exists to fix
v1.7.x had risk_taxonomy.json keyed on mechanism only (e.g. "egfr_combo_pdac": {...narrative about EGFR combo in PDAC...}). The risk_lookup.py would attach this risk to ANY trial whose mechanism matched, regardless of patient cancer type. Result: a CRC patient on a Calderasib (KRAS G12C) trial for advanced solid tumors got the risk note "在 PDAC 中常作为 KRAS G12D / G12C 抑制剂的联合伙伴 (克服 EGFR-feedback)" — wrong cancer context, sounds clinically authoritative, but doesn't apply to CRC.
v2 rule: every risk annotation MUST be triple-grounded in (mechanism, cancer_type, patient_state). No risk template gets emitted unless all three are checked.
Inputs
{
"trial": {
"id": "NCT...",
"title": "...",
"phases": [...],
"interventions": ["sotorasib", "cetuximab"],
"mechanism_of_action": "KRAS G12C inhibitor + anti-EGFR mAb"
},
"patient": { }
}
Output
{
"trial_id": "NCT...",
"risks": [
{
"key": "kras_g12c_inhibitor_class",
"mechanism": "KRAS G12C covalent inhibitor",
"cancer_context": "CRC (patient's tumor type)",
"risk_level": "moderate",
"narrative": [
"KRAS G12C inhibitor monotherapy in CRC has substantially lower ORR than NSCLC (CRC ~10%, NSCLC ~30-40%) due to adaptive EGFR feedback unique to CRC biology",
"Combination with anti-EGFR mAb (cetuximab/panitumumab) restores ORR to ~30% per CodeBreaK 300 (sotorasib + panitumumab vs investigator choice)",
"If trial is monotherapy in CRC cohort, expect lower ORR; if trial includes anti-EGFR combo arm, mention this paradigm to patient"
],
"applies_because": "Trial intervention includes KRAS G12C inhibitor; patient cancer is CRC"
}
]
}
Process
1. Identify the trial's mechanism(s) of action from interventions
2. For each mechanism, evaluate (mechanism, patient.cancer_type, patient.prior_therapies):
- Look up rules/{mechanism}-{cancer}.md if it exists (most specific)
- Otherwise rules/{mechanism}.md with cancer-specific notes inline
- Otherwise emit narrative from training knowledge with explicit grounding
3. For each risk, write narrative that names the cancer context explicitly
4. Tag risk_level: low | moderate | high | high_uncertainty
5. Filter out any risk not actually applicable to this (mechanism, cancer) pair
Risk-level taxonomy
| Level | Meaning |
|---|
low | Well-characterized risk, established management protocol |
moderate | Predictable AE pattern, may require dose modification or supportive care |
high | Significant patient-specific concern (comorbidity interaction, prior failure mode) |
high_uncertainty | First-in-class / phase 1 dose escalation / limited published data — outcomes unpredictable |
Mandatory grounding check
Before emitting any risk narrative, verify:
- ✅ Does this risk apply to the patient's specific cancer type? (Reject if narrative is from a different cancer context — see PDAC-leak-to-CRC example)
- ✅ Does this risk consider the patient's prior therapy load? (e.g. don't warn about "PD-1 toxicity" for a patient who's been on camrelizumab for a year without irAE)
- ✅ Is the narrative specific enough to be actionable? (Reject "watch for AEs" — name the AE, name the management)
If a risk doesn't pass these checks, omit it.
Reference: cancer-specific patterns
