Hcc Meta Ai

HCC Meta-Analysis AI Skill

Overview

This skill implements a standardized, AI-assisted meta-analysis workflow for comparing treatment efficacy in hepatocellular carcinoma (HCC). It follows a six-element prompt architecture (Role–Objective–Skills–Constraints–Workflow–Output) validated across two independent clinical questions with three large language models.

The workflow covers the full meta-analysis pipeline:

1. Literature screening (fuzzy → precise → supplementation)
2. Full-text eligibility determination
3. Newcastle–Ottawa Scale (NOS) quality appraisal
4. Structured baseline data extraction
5. Clinical outcome data extraction

Validation: Benchmarked against dual-independent manual review across two HCC projects with GPT-5.2, GPT-4o, and DeepSeek V3.1. Best-performing models achieved >95% accuracy at most stages with >80% total time reduction.

When to Use

• Conducting a meta-analysis comparing two HCC treatments
• Need rapid evidence synthesis for HCC therapeutic comparison
• Want to standardize the AI-assisted review process across projects
• Teaching or replicating the validated AI meta-analysis methodology

Six-Element Prompt Framework

Every prompt in this workflow follows a standardized six-element structure:

Workflow Stages

Stage 1: Literature Screening

Three sub-stages with escalating cognitive demands:

1. Fuzzy Screening — Title-level keyword matching

   • Batch 100 records per query
   • Check for presence of target disease + intervention terms
   • Template: scripts/01-fuzzy-screening.md

2. Precise Screening — Abstract-level multi-constraint reasoning

   • Evaluate title + abstract against inclusion criteria
   • Requires comparison of two interventions with reported outcomes
   • Template: scripts/02-precise-screening.md

3. Literature Supplementation — Citation tracing from full texts

   • Upload PDFs of included studies
   • Extract reference lists; flag potentially missed eligible studies
   • Template: scripts/03-literature-supplement.md

Stage 2: Full-Text Eligibility Determination

Full-text comprehension and multi-criteria decision-making:

• Study population (primary HCC confirmed)
• Direct comparison of two target treatments
• ≥1 clinical outcome reported
• Original study design (prospective/retrospective cohort, case-control, RCT)
• Data separable by treatment group
• Sample size ≥10
• Template: scripts/04-eligibility.md

Stage 3: NOS Quality Appraisal

Structured scoring across three domains:

• Selection (4 items, max 4 pts): representativeness, non-exposed cohort, exposure ascertainment, outcome absent at baseline
• Comparability (1 item, max 2 pts): confounder control
• Outcome (3 items, max 3 pts): assessment objectivity, follow-up duration, follow-up completeness
• Template: scripts/05-nos-scoring.md

Stage 4: Baseline Data Extraction

Structured extraction of study characteristics and patient demographics:

• Core fields: author, year, design, treatment groups, sample size, age, sex, tumor characteristics, follow-up, laboratory values
• Topic-specific extensions (see references)
• Template: scripts/06-baseline-extraction.md

Stage 5: Outcome Data Extraction

Structured extraction of clinical endpoints:

• 1–5 year overall survival (OS), recurrence-free survival (RFS)
• Local tumor progression, distant recurrence, technical success
• Complication rates and subtypes
• Template: scripts/07-outcome-extraction.md

Disease-Agnostic Design

The six-element scaffold is independent of specific diseases. To adapt this workflow for a different disease:

1. Replace disease terminology in references/terminology.json
2. Swap intervention names in each prompt template
3. Adjust topic-specific extraction fields in baseline/outcome templates
4. Keep the core logical structure intact

For HCC specifically, the pre-configured terminology and extraction fields in references/ are ready to use.

Platform

A web-based platform for deploying and sharing this workflow is available at: http://8.149.142.6/metaexp/

References

• Full methodology: see accompanying manuscript
• Supplementary prompt details: S1 (Project 1 prompts), S2 (Project 2 prompts)
• Terminology database: references/terminology.json