Systematic Review Screening Assistant

Use meta-catalog, meta-search, and agentic-search for PRISMA-style initial screening

Scenario

Researchers in medicine, life sciences, materials, and related fields need PRISMA-style initial screening to select candidate papers from a large literature pool.

Estimated calls

~30-80 API calls per screening task

Tools

meta-catalogmeta-searchagentic-searchcontent

Pipeline

meta-catalog→ 确认可筛字段→ meta-search 广撒网→ agentic-search 精筛→ PRISMA 流程图

Input example

Screening criteria:
- population: patients with Alzheimer disease
- intervention: blood-based biomarkers
- year: 2020-2025
- include clinical studies only

Output example

Screening table:
- paper_id, title, year, include/exclude, reason, evidence quote
- PRISMA counts: retrieved, screened, included, excluded

Agent Prompt example

You are a systematic review screening assistant. Use structured filters to build the candidate pool, then semantic search and source verification to judge inclusion criteria.

Implementation steps

Step 1: Set up the environment

Configure token and dependencies

!pip install httpx pandas
import os
os.environ["SCIVERSE_API_TOKEN"] = "sv-your-token-here"  # 替换为你的真实值

Step 2: Query available screening fields

Use meta-catalog to identify useful filter fields

import os
import asyncio
import httpx

BASE = "https://api.sciverse.space"
TOKEN = os.environ["SCIVERSE_API_TOKEN"]
HEADERS = {"Authorization": f"Bearer {TOKEN}"}

async def get_catalog():
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.get(f"{BASE}/meta-catalog", headers=HEADERS)
        resp.raise_for_status()
        return resp.json()["fields"]

fields = await get_catalog()
for f in fields:
    print(f"{f['name']} ({f.get('type','')}): operators={f.get('operators', [])}")

Step 3: Run broad retrieval

Use meta-search to build the candidate set

import pandas as pd

async def broad_search(query: str, year_from: int, year_to: int, page_size: int = 100):
    """\u5e7f\u6492\u7f51: \u6309\u5e74\u4efd\u8303\u56f4\u68c0\u7d22\u6240\u6709\u5019\u9009\u6587\u732e"""
    all_results = []
    page = 1
    while True:
        async with httpx.AsyncClient(timeout=30) as client:
            resp = await client.post(
                f"{BASE}/meta-search", headers=HEADERS,
                json={
                    "query": query,
                    "filters": [
                        {"field": "publication_published_year", "operator": "FILTER_OP_GTE", "value": year_from},
                        {"field": "publication_published_year", "operator": "FILTER_OP_LTE", "value": year_to},
                    ],
                    "page": page, "page_size": page_size
                }
            )
            resp.raise_for_status()
            data = resp.json()
            all_results.extend((data.get("results") or []))
            if len(all_results) >= data.get("total_count", 0) or len((data.get("results") or [])) < page_size:
                break
            page += 1
    return all_results, data.get("total_count", 0)

INCLUSION_QUERY = "CAR-T cell therapy solid tumor clinical trial"
results, total = await broad_search(INCLUSION_QUERY, 2019, 2024)
print(f"Identification: {total} records found")

Step 4: Semantic screening and inclusion judgment

Use agentic-search and content to judge inclusion criteria with evidence

async def semantic_screen(candidates: list[dict], query: str, top_k: int = 100):
    """\u8bed\u4e49\u7cbe\u7b5b: \u7528 agentic-search \u5bf9\u5019\u9009\u6587\u732e\u8bc4\u5206"""
    candidate_ids = {r["doc_id"] for r in candidates if r.get("doc_id")}
    async with httpx.AsyncClient(timeout=30) as client:
        resp = await client.post(
            f"{BASE}/agentic-search", headers=HEADERS,
            json={"query": query, "top_k": top_k}
        )
        resp.raise_for_status()
        hits = (resp.json().get("hits") or [])
    return [h for h in hits if h.get("doc_id") in candidate_ids]

hits = await semantic_screen(
    results,
    "CAR-T cell therapy clinical trial solid tumor patients outcomes"
)

# \u6309\u76f8\u5173\u6027\u5206\u6570\u7b5b\u9009
# screening 现在是 identification 候选池的子集
screened = [h for h in hits if h["score"] >= 0.7]
print(f"Screening: {len(screened)} records (score >= 0.7)")

# \u8f93\u51fa PRISMA \u6d41\u7a0b\u6570\u636e
prisma = {
    "identification": total,
    "screening": len(screened),
    "included": len([h for h in screened if h["score"] >= 0.85])
}
print(f"\
PRISMA Flow: {prisma}")

# \u5bfc\u51fa CSV
df = pd.DataFrame(screened)
df.to_csv("screened_papers.csv", index=False)
print("Exported to screened_papers.csv")

Notes

Keep inclusion and exclusion criteria explicit.
Store reasons for every excluded paper.
Use structured filters for the first pass and semantic evidence for the second pass.
Do not make final clinical conclusions without human review.
Record PRISMA-style counts for auditability.

FAQ

是否能替代人工系统综述?

不能，只适合自动化初筛和候选集整理，最终纳入仍需人工判断。

如何控制筛选条件?

用 meta-catalog 确认字段，再用 meta-search 和 agentic-search 逐步筛选。

输出应包含什么?

建议包含候选数量、纳入理由、排除理由、关键字段和可复核证据。

如何降低漏筛?

先用宽条件构建候选池，再逐步加入语义筛选和人工复核。

Next steps

Need an API key?

Create one in Console > Tokens.The same API key works for enabled Sciverse, DianShi, and Skills capabilities, with starter quota available according to account permissions.

Open console