PRAIM is not Denmark. Germany kept two human readers and added a machine that can force a second look after a human-normal call. Detection rose at national implementation scale. Do not sell this as retired double reading; the paper’s 56.7% automation scenario is hypothetical.
AI value · Model-based breast cancer detection rate
Verified
6.70 per 1,000 in the AI-supported group; +17.6% (95% CI +5.7% to +30.8%) after overlap weighting
Observational; radiologists chose the viewer. Authors used propensity-score overlap weighting because some readers preferentially used AI on normal-tagged exams. Not a reader-replacement deployment.
Before
Independently double-read four-view mammograms without AI predictions. → Confirms or dismisses suspicion and issues recall.
After
Tags confident-normal cases and fires a localization safety net when humans call suspicious cases normal. → Reads with normal tags visible and must accept or reject AI localization after a safety-net prompt.
Human boundary
Radiologists choose the viewer per exam. The safety net can force reconsideration; it cannot recall a woman.
Why it matters
Two unaided human readers plus consensus are not the only safe model for a national screening programme.
How the work changed
Before
How the work ran before the change.
Step 1 of 2
Two radiologists
Independently double-read four-view mammograms without AI predictions.
ControlBinding national double-reading guideline
Step 2 of 2
Consensus conference
Confirms or dismisses suspicion and issues recall.
ControlAt least the two readers plus a head radiologist
What changed
Two unaided human readers plus consensus are not the only safe model for a national screening programme.
Decision rightAI acts within a human backstop
After
How the same work runs now.
Step 1 of 2
Vara MG AI
Tags confident-normal cases and fires a localization safety net when humans call suspicious cases normal.
ControlCE-marked device; live vendor monitoring; radiologist may ignore the AI viewer
Step 2 of 2
Radiologist (AI viewer)
Reads with normal tags visible and must accept or reject AI localization after a safety-net prompt.
ControlFinal recall remains human; 204 cancers in the AI group were diagnosed after accepted safety-net prompts
Process model built from the published workflow evidence for German organised mammography screening programme (PRAIM implementation study; 12 sites). Every step, actor, and control appears in full below.Every step, actor, and control
Exception path
Readers can stay on the non-AI viewer. Examinations tagged normal by AI still produced 20 cancers in the AI group after human consensus, so the normal tag is not an autonomous discharge.
Decision authority
Radiologists choose the viewer per exam. The safety net can force reconsideration; it cannot recall a woman.
Before
#
Actor
Action
Control
01
Two radiologists
Independently double-read four-view mammograms without AI predictions.
Binding national double-reading guideline
02
Consensus conference
Confirms or dismisses suspicion and issues recall.
At least the two readers plus a head radiologist
After
#
Actor
Action
Control
01
Vara MG AI
Tags confident-normal cases and fires a localization safety net when humans call suspicious cases normal.
CE-marked device; live vendor monitoring; radiologist may ignore the AI viewer
02
Radiologist (AI viewer)
Reads with normal tags visible and must accept or reject AI localization after a safety-net prompt.
Final recall remains human; 204 cancers in the AI group were diagnosed after accepted safety-net prompts
Work that left the path
Some reread time on AI-normal exams (median 16 s versus 30 s unclassified in the AI group)
Not a removed second reader; that scenario is a post-hoc analysis, not the deployed protocol
Human role before
Two unaided readers plus consensus owned every interpretation.
Human role after
Readers who opt into the AI viewer work a triaged list and a forced second look on AI-suspicious cases they had called normal. Recall remains human. Double reading is not removed.
AI role
Confident-normal triaging and a post-read safety net. It does not replace a reader in this study design.
Outcomes
Model-based breast cancer detection rate
Verified
5.70 per 1,000 in the contemporaneous non-AI viewer group→6.70 per 1,000 in the AI-supported group; +17.6% (95% CI +5.7% to +30.8%) after overlap weighting
2021-07-01 to 2023-02-23 · 461,818 women analysed; 260,739 AI group; 201,079 control; 119 radiologists; 12 sites
Observational; radiologists chose the viewer. Authors used propensity-score overlap weighting because some readers preferentially used AI on normal-tagged exams. Not a reader-replacement deployment.
What leaders can reuse
Anti-pattern
Reporting the fictitious ‘do not read AI-normals’ scenario as the deployed workflow.
Questions
01If we cannot drop a second reviewer, will we still force a machine interrupt after a human-normal call?
02How will we stop readers from using the AI path only on easy exams?
Portability conditions
A double-reading legal framework you cannot immediately change
Willingness to insert a post-read interrupt rather than drop a reader
Confounding control if clinicians self-select into the AI path
Reputation risk
medium if described as autonomous screening. It is AI-supported double reading with voluntary uptake.
Evidence and authority
What the public record supports.
Current · updated
2 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-09-06 · stable ID 41cba9f08878beb9
