shield_lock
金大哥
← 返回 Skill 列表
extension
分类: 效率与办公无需 API Key

sr-interpretation

Structured interpretation of SR results per Cochrane Handbook Ch 15. Use when the user needs to interpret meta-analysis results, compute NNT, re-express SMDs, assess applicability, or draw conclusions for practice and research. Trigger on "interpret", "NNT", "number needed to treat", "applicability", "SMD interpretation", "clinical significance", "implications for practice", "implications for research", "draw conclusions", "certainty interpretation", "absolute risk", "relative risk", "MID". Use ONLY after synthesis and GRADE are complete.

person作者: TashanworldhubOpenAPI
open_in_new仓库download下载资源包

Interpretation and Drawing Conclusions (Cochrane Handbook Ch 15)

Your Role

Structure the interpretation of SR findings: applicability assessment, NNT computation, SMD re-expression, correct use of CIs and P values, and framing implications for practice and research.

Prerequisites

  • Pooled effect estimates (from sr-synthesis)
  • GRADE certainty ratings (from sr-grade)
  • Study characteristics (from sr-extraction)

Workflow

Step 1: Assess Applicability (Ch 15.2)

CHECK each dimension:

Biological variation: Are the study populations similar to the target population? (age, sex, disease severity, comorbidities)

Contextual variation: Are the settings similar? (ED, ICU, pre-hospital — resource availability, staffing, protocols)

Variation in adherence: Would real-world adherence differ from trial conditions?

Values and preferences: Would patients in the target setting value outcomes the same way?

For each, document: "The evidence is [directly applicable / somewhat indirect / very indirect] for [dimension]."

Step 2: Compute Absolute Effects and NNT (Ch 15.4)

If the MA used RR or OR:

python3 scripts/nnt_calculator.py --effect RR --value 0.75 --ci-lower 0.65 --ci-upper 0.85 --acr 0.30

GENERATE results for LOW, MEDIUM, and HIGH baseline risk:

Assumed Comparator Risk (ACR): 150 per 1000 (low)
Absolute effect: 37 fewer per 1000 (95% CI: 23 fewer to 49 fewer)
NNTB: 27

Assumed Comparator Risk (ACR): 300 per 1000 (medium)
Absolute effect: 75 fewer per 1000 (95% CI: 45 fewer to 105 fewer)
NNTB: 14

Assumed Comparator Risk (ACR): 500 per 1000 (high)
Absolute effect: 125 fewer per 1000 (95% CI: 75 fewer to 175 fewer)
NNTB: 8

LABEL correctly: NNTB = number needed to treat for an additional BENEFICIAL outcome. Use NNTH for harmful outcomes.

Step 3: Interpret Continuous Outcomes (Ch 15.5)

Same scale (e.g., both studies use APACHE II): report MD in original units. "If the MD is larger than the MID (minimal important difference), the effect is clinically meaningful."

Different scales (e.g., pain on VAS and NRS): report SMD. RE-EXPRESS SMD in natural units:

python3 scripts/smd_interpreter.py --smd 0.45 --pooled-sd 8.5

"An SMD of 0.45 corresponds to a MD of 3.8 points on the APACHE II scale (pooled SD from included studies: 8.5)."

If MID is known: compare SMD to MID/SD ratio. SMD > MID/SD → clinically meaningful.

Step 4: Correctly Handle CIs and P Values (Ch 15.3)

Do NOT use "statistically significant" or "non-significant" (MECIR C72 mandatory).

INSTEAD describe:

  • "The 95% CI excludes the null (RR 0.75, 95% CI 0.65 to 0.85)"
  • "The 95% CI includes the null (RR 0.92, 95% CI 0.81 to 1.04)"
  • "The CI is wide, reflecting imprecision (RR 0.70, 95% CI 0.30 to 1.60)"

Also check the PREDICTION INTERVAL (from sr-synthesis): does the PI cross the null even if the CI does not? If yes: "The average effect is beneficial, but the effect varies across settings and may be null or harmful in some."

Step 5: Generate Implications (Ch 15.6)

Implications for Practice

Structure as:

  • Who this evidence applies to (and who it does not)
  • What intervention, compared to what
  • What is the balance of benefits and harms
  • Certainty of the evidence (from GRADE)
  • Values and preferences that might influence decisions

"Do NOT make recommendations — only highlight actions consistent with particular values."

Implications for Research

Structure as:

  • What specific research is needed
  • Why (precise gap: population, intervention, comparator, outcome, design)
  • How (suggested study design, sample size, follow-up)
  • Priority (high/medium/low)

"If the evidence is sparse, say: 'Further research is likely to change the estimate.'"

Scripts

scripts/nnt_calculator.py

Computes NNT/NNTB/NNTH from RR/OR with CI across multiple ACRs. Uses Altman (1998) method for NNT CIs. Usage: python3 nnt_calculator.py --effect RR --value 0.75 --ci-lower 0.65 --ci-upper 0.85 --acr 0.1,0.3,0.5

scripts/smd_interpreter.py

Re-expresses SMD in natural units or MID units. Usage: python3 smd_interpreter.py --smd 0.45 --pooled-sd 8.5 or --mid 5.0

scripts/applicability_checklist.py

Structured indirectness assessment across 4 dimensions. Usage: python3 applicability_checklist.py --pico sepsis,vasopressors,mortality,icu

Assets

assets/nnt-interpretation-guide.md

NNTB vs NNTH labeling, time-frame requirement, event-rate dependence.

assets/implications-template.md

Templates for "Implications for practice" and "Implications for research" sections.

assets/mid-reference-em.md

Published MIDs for common EM/CC outcomes: mortality (RR 0.90), vent days (MD 2), LOS (MD 1 day), pain VAS (MD 1.5 cm), SOFA (MD 1).

Guardrails

  1. "Do NOT make recommendations — Ch 15.6: 'Review authors should avoid specific recommendations.'"
  2. "Do NOT use 'statistically significant' or 'non-significant' — MECIR C72 mandatory."
  3. "NNT must specify: intervention, comparator, outcome, time-frame, and ACR."
  4. "For SMD <0.2, even if statistically significant, clinical importance is questionable."
  5. "Prediction interval crossing the null is important even if the CI does not."
  6. "Absolute effects depend on baseline risk — report across a RANGE of risks, not just one."
  7. "The so-called 'number needed to harm' (NNH) is misleading — use NNTH instead."

Handoff

→ sr-writing (discussion + conclusions sections)