科研 Skill

多模态大模型临床评测设计

mllm-eval

Audits how an LLM or multimodal LLM task is evaluated, covering reference standard and hallucination. Use when reviewing an evaluation harness for a clinical AI model.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

MLLM-Eval Skill

Purpose

This skill makes an LLM / MLLM clinical evaluation defensible: a real adjudicated reference standard, faithfulness measured not assumed, clinical-efficacy metrics beyond n-gram overlap, a pretraining- contamination check, prompt-sensitivity disclosed, and a reader study where text is generated. It is model-agnostic — every check applies to a closed API and to open weights — and read-only (an advisory design/audit skill): it audits the evaluation design and specifies and routes the clinical-efficacy metrics (RadGraph-F1 / CheXbert-F1 via their published extractors) rather than running the model or computing the metrics itself.

It is the LLM/MLLM evaluation-design counterpart in the lane — an auditor that hands the specified metrics to their extractors and /analyze-stats, parallel to how /model-assessment audits an imaging model's design and computes its metrics. The reviewer-side audit of a finished manuscript uses the mllm_evaluation.md (ME0–ME8) probe via /self-review and /peer-review; this skill is the author-side harness design. It routes the reader study to /design-ai-benchmarking, the sizing to /calc-sample-size, and TRIPOD-LLM / MI-CLEAR-LLM compliance to /check-reporting.

When to use

  • You are designing or auditing an evaluation of an LLM/MLLM on a clinical task and want it to cover the axes a reviewer will check (reference standard, faithfulness, contamination, prompt sensitivity, reader study).

When NOT to use

  • AI-vs-human-expert benchmark with a rated rubric → /design-ai-benchmarking.
  • Imaging prediction/segmentation model → /model-assessment.
  • Image-to-image generative model → the image_synthesis probe.
  • Training / serving the LLM → out of scope.
  • Item-level TRIPOD-LLM / MI-CLEAR-LLM audit of a finished manuscript → /check-reporting.

Workflow

Phase 1 — Pin the task, model, comparator, decoding (ME0)

State the task (report generation / VQA / extraction-classification), the exact model + version/date (closed API or open-weights id), the decoding settings (temperature, seed, max tokens), and what the outputs are scored against.

Phase 2 — Reference standard + metrics (ME1, ME2)

Require an adjudicated expert reference (not a single unverified report or a model-derived label). For report generation, report a clinical-efficacy metric — RadGraph-F1 (Jain et al., NeurIPS 2021) or CheXbert-F1 (Smit et al., 2020), or the composite RadCliQ (Yu et al., Patterns 2023) — alongside any BLEU/ROUGE, with CIs. For VQA/classification, state the answer-matching rule and report per-class sensitivity/specificity (or precision/recall/F1) and PPV at the real prevalence, with CIs; accuracy only alongside them — at 2% prevalence, answering "negative" every time scores 98% accuracy.

Phase 3 — Faithfulness + contamination (ME3, ME4)

Add an atomic-fact faithfulness measure + a false-premise / abstention probe (MedVH, Med-HALT) — report a hallucination rate, not just accuracy. For any public benchmark (VQA-RAD, SLAKE, MIMIC-CXR- derived, MedQA), add a contamination statement: training cutoff vs benchmark release, a held-out / post-cutoff set, or a contamination probe.

Phase 4 — Prompt sensitivity + reader study (ME5, ME7)

Disclose the exact prompt(s), temperature/seed, ≥ 3 runs with variance, and a prompt-robustness check. For a deployment/utility claim, design a blinded reader study with an error taxonomy (route the rubric/IRR to /design-ai-benchmarking, ICC/κ to /analyze-stats, sizing to /calc-sample-size).

Phase 5 — Gate the plan (deterministic)

Declare the axes in eval_manifest.json (copy templates/eval_manifest.json; fields and allowed values in references/eval_manifest_schema.md), then:

python3 scripts/check_mllm_eval_completeness.py \
  --manifest eval_manifest.json --strict

Use "none" for an axis not done and "other:<description>" for a method not listed; any other value exits 2. --plan plan.md --task report_generation|vqa|classification still runs the older keyword check on prose (see Known limits). NGRAM_ONLY / FAITHFULNESS_MISSING / REFERENCE_STANDARD_MISSING / CONTAMINATION_UNADDRESSED / READER_STUDY_MISSING must be resolved. A classification manifest always reports CLASSIFICATION_METRICS_NOT_ASSESSED (Minor): the manifest has no metric field for it, so check per-class sensitivity/specificity and PPV at the real prevalence by eye.

Phase 6 — Hand off

Methods/Results → /write-paper; compliance (TRIPOD-LLM / MI-CLEAR-LLM) → /check-reporting; reviewer audit → /self-review (loads ME0–ME8).

Anti-Hallucination

  • Never fabricate model outputs, reference labels, or metric scores. Compute only what the supplied outputs allow; if a reference standard or outputs are missing, say so and stop.
  • Never report n-gram overlap (BLEU/ROUGE) as clinical correctness — pair it with a clinical-efficacy metric, and flag the n-gram score for what it is.
  • Never claim "no contamination" without a stated check when a public benchmark is used.
  • If a metric (RadGraph-F1 / CheXbert-F1) or its extractor is uncertain, flag [VERIFY] and ask rather than inventing a number.

Deterministic gate

scripts/check_mllm_eval_completeness.py — task-aware presence gate on the evaluation plan (stdlib, network-free). Reproducible challenge: bash scripts/mllm_eval_completeness_challenge/verify.sh.

Known limits. Manifest mode checks what is declared, not that the work was done; keep the manifest in step with the Methods. It also does not check that a declared metric suits the task: report-generation metrics (BLEU, RadGraph F1) declared for a VQA task are accepted without comment. Prose mode (--plan) checks that a term is present; it does not read negation or sense. "No human evaluation was performed" or "hallucination was not assessed" still counts as covering that axis, and a word used in another sense still clears it: "green" anywhere clears the clinical-metric check, "unsupported" clears faithfulness, "data leakage" from a patient split clears contamination, and "ground truth" clears the reference standard whatever it refers to. Read each cleared axis in the plan yourself before treating an exit 0 as clean.

Boundaries

mllm-eval (this skill: harness design + completeness gate, model-agnostic)
  ├─ design-ai-benchmarking (reader-study rubric / IRR)
  ├─ calc-sample-size (reader + case sizing)
  ├─ write-paper + check-reporting (TRIPOD-LLM / MI-CLEAR-LLM)
  └─ self-review / peer-review (ME0–ME8 reviewer probe)

Reference Files

  • references/evaluation_axes.md — the why behind the ME2–ME7 axes: clinical-efficacy metrics beyond n-gram overlap (e.g. RadGraph-F1 / CheXbert-F1 vs BLEU/ROUGE), faithfulness & hallucination, pretraining/benchmark contamination, prompt-sensitivity & determinism, answer-matching, and the reader study — each mapped to its gate verdict. Load on demand during Phases 2–4.

示例提示词

帮我设计一套评测多模态大模型生成放射报告质量的方案。

所属插件

临床医学与公共卫生面向临床与公卫研究:研究设计与方案、样本量、队列与 Meta 分析、报告规范核查、医学影像 AI 评估、流行病建模与卫生经济学评价、医学论文写作投稿。

安装该插件会一并获得同领域的 58 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

多模态大模型临床评测设计能做什么?

Audits how an LLM or multimodal LLM task is evaluated, covering reference standard and hallucination. Use when reviewing an evaluation harness for a clinical AI model。

如何安装多模态大模型临床评测设计?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「多模态大模型临床评测设计」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

多模态大模型临床评测设计需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

多模态大模型临床评测设计经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。