biomed-evidence-synthesizer

Synthesize biomedical evidence across multiple studies, including randomized trials, observational cohorts, case reports, animal models, in vitro experiments, longitudinal studies, and meta-analyses.

SKILL.md

GitHub

Biomedical Evidence Synthesizer

Integrate multiple biomedical studies into a balanced conclusion. Use this when the user has several papers, mixed evidence, conflicting results, or a broad question that requires weighing clinical, observational, animal, and mechanistic data together.

Step 0: Frame the Evidence Question

Write the question as:

In [population/context], does [exposure/intervention/mechanism] affect
[endpoint/outcome], compared with [control/comparator], over [time scale]?

If needed, split one broad question into mechanistic, efficacy, safety, and translation subquestions.

Step 1: Build the Evidence Table

Group studies by type:

  • Systematic reviews and meta-analyses.
  • Randomized controlled trials.
  • Nonrandomized interventional studies.
  • Prospective cohorts.
  • Retrospective cohorts and case-control studies.
  • Case reports and case series.
  • Animal studies.
  • Ex vivo, organoid, in vitro, biochemical, and computational studies.
  • Trial registry records and regulatory reviews.

For each study, extract:

  • Population/model.
  • Exposure/intervention and comparator.
  • Dose, route, timing, duration, and follow-up.
  • Endpoint type: clinical, surrogate, biomarker, mechanistic, safety.
  • Effect direction and magnitude.
  • Statistical uncertainty.
  • Main limitations.

Step 2: Compare Studies

Check whether studies are actually asking the same question. Differences that often explain disagreement:

  • Disease subtype or stage.
  • Age, sex, comorbidities, baseline risk, and concomitant therapies.
  • Species, strain, tissue, cell line, or model system.
  • Dose, formulation, route, timing, and treatment duration.
  • Acute versus chronic exposure.
  • Primary versus secondary endpoints.
  • Surrogate versus clinical endpoints.
  • Adjustment for confounding.
  • Follow-up length.
  • Publication bias, selective outcome reporting, and sponsor effects.

Step 3: Grade the Body of Evidence

Assess:

  • Consistency: are effects in the same direction?
  • Replication: independent groups, populations, and methods.
  • Precision: sample size, uncertainty intervals, event counts.
  • Directness: does evidence answer the actual clinical or mechanistic question?
  • Bias risk: randomization, blinding, confounding, attrition, selective reporting.
  • Dose-response and temporal relationship.
  • Biological plausibility without over-weighting mechanism.
  • Safety and adverse event evidence.

Use these conclusion labels:

  • Clinically established.
  • Clinically supported but incomplete.
  • Promising but preliminary.
  • Mechanistically plausible but clinically unproven.
  • Mixed or inconclusive.
  • Unsupported.
  • Contradicted by stronger evidence.

Step 4: Resolve Conflicts

When studies disagree, do not average them rhetorically. Identify whether the conflict is likely due to:

  • Study quality.
  • Different populations or models.
  • Different endpoints.
  • Different doses or timing.
  • Chance, underpowering, or multiple testing.
  • Confounding or bias.
  • Real biological heterogeneity.

Say which explanation is most plausible and why.

Step 5: Output

Use this structure:

# Evidence Synthesis: [Question]

## Bottom Line
[One cautious paragraph with evidence grade.]

## Evidence by Study Type
[Short summary grouped by evidence type.]

## Cross-Study Comparison
| Study | Type | Population/Model | Exposure | Endpoint | Result | Key Limitation |
|------|------|------------------|----------|----------|--------|----------------|

## Areas of Agreement
[Findings replicated or directionally consistent.]

## Areas of Conflict
[Contradictions and plausible explanations.]

## Human Clinical Relevance
[Whether evidence reaches human clinical endpoints.]

## Safety Evidence
[Known adverse events, missing safety data, follow-up limitations.]

## Most Plausible Interpretation
[Reasoned synthesis, not a vote count.]

## What Would Change the Conclusion
[Specific future study designs, endpoints, or missing data.]

Standards

  • Weight stronger designs more heavily, but do not ignore consistent lower-level evidence.
  • Do not equate mechanistic coherence with clinical proof.
  • Do not bury heterogeneity; explain it.
  • Treat unpublished completed trials and terminated trials as important signals when relevant.
  • Keep causality claims proportional to design quality.