Skip to content
No Vape

A vaping review used AI screening: find what the tool actually decided

Distinguish ordering records for people from excluding them automatically, and keep the checking process attached to the review’s coverage claim.

No Vape · Updated

Open book on a cafe table
Illustrative image · Lucille Emi Oh / Unsplash

Locate the tool within the review process

A review has several stages, including finding records, deciding eligibility and interpreting included studies. Using AI somewhere does not tell you which stage changed. Mark the named task and the point at which a record could stop receiving further attention. Do not confuse selecting research papers with selecting treatment for an individual.

Cochrane’s original study-selection chapter describes several automation approaches, including classifiers operating across reviews and systems learning within an individual review. It discusses workload and sensitivity trade-offs. We use the distinction between tasks, without repeating performance percentages or endorsing a tool for all vaping evidence.

Cochrane: Cochrane Handbook Chapter4, section4.6.6.2 — Automating the selection process

Make a decision-and-checking flow

For each stage, note input, tool action, human action and what happens to records the tool does not prioritize or retain. Ordering a queue is different from a final exclusion. If a stopping rule is reported, preserve it. If the report does not say what happened to the remainder, leave that part of the flow unresolved.

Consider a fictional review that ranks titles so people read likely relevant records first. Another imaginary workflow automatically removes records given a particular classification. Both could be described as AI-assisted, yet the route to exclusion differs. This exercise contains no actual review or claim that either workflow necessarily missed a study.

Ask how the task was checked in this setting

Find any evaluation, human verification or safeguards the report describes for the actual selection task. An excellent score on a different dataset does not document this review’s decisions. Keep the meaning of the labels clear: a trial-design classifier may not answer whether a trial meets this review’s population and outcome criteria.

Also distinguish numbers of records from numbers of independent studies. Several reports can describe one study even when software treats them as separate records initially. Automation does not remove the need to understand that grouping. Do not reconstruct a new evidence total by adding every item appearing in a search display.

Keep coverage and conclusions as separate questions

A recent search date tells you when searching occurred, not how every retrieved record was screened. Conversely, a well-described screening tool does not show that every possible information source was searched. Your reading note should name the automated decision, checks and unresolved limitation rather than replacing several methodological questions with the label AI.

This is a workflow-reading exercise, not a formal audit or a replacement systematic review. For clinical interpretation seek suitable expertise with the complete evidence. You can finish by explaining which decisions remained with people and which did not, while leaving effectiveness of any quitting product to the intervention evidence that actually addresses it.

Sources and scope

Sources inform the guidance. The worksheets and examples are No Vape editorial suggestions, not tested treatments.

  1. Cochrane — Cochrane Handbook Chapter4, section4.6.6.2 — Automating the selection process
  2. No Vape — No Vape — app features and store links

Written with AI assistance and checked against the cited sources. No clinical review is claimed. This guide covers practical planning, not individual medical treatment. How we research and write