Skip to main content
AIH Lab
Sign in
LibraryFamily 15 · Emergency & Trauma

Implementation of machine learning software on the radiology worklist decreases scan view delay for intracranial haemorrhage on CT (Brain Sciences 2021)

Emerging evidence

Retrospective comparison of scan view delay — time from study completion to a radiologist opening it — between scans flagged positive for intracranial haemorrhage by the algorithm and non-flagged scans.

ReviewedPubMed · PMID:34201775
What this means:
A curator has read the source and confirmed our reading of it — the classifications that let this record be compared with others — along with the AI product and the clinical use. That is a check of our work, not a judgement on the AI.
What would change it:
The pair enters the AIH factory, where clinical, technical and governance findings are worked section by section.
  1. Submitted
  2. 2Curated
  3. 3Maturing
  4. 4Validated
  5. 5Scaled
90
Other
8,723
Participants
United States
Country
Retrospective single-centre cohort
Study

Plain-language summary

AI

Choose an audience and generate a tailored summary on demand. The AI uses only what is on this page.

Clinical context
Clinical problem
Detect a condition
Point of care
Triage
Nature of AI output
A prioritised list
Clinical specialty
Radiology
Care setting
hospital_emergency
Patient population
Patients undergoing head CT across outpatient, inpatient and emergency settings at a single academic medical centre.
Intended use
Retrospective comparison of scan view delay — time from study completion to a radiologist opening it — between scans flagged positive for intracranial haemorrhage by the algorithm and non-flagged scans.
Full study description

8,723 head CTs (1,829 flagged positive, 6,894 not flagged). Worklist prioritisation reduced scan view delay by 604 minutes for outpatients (90%, p<0.0001) and 38 minutes for inpatients (10%, p=0.002). The emergency-setting reduction of 12.5 minutes (15%) was NOT statistically significant (p=0.37) — recorded because the workflow benefit is strongest where baseline delay is longest, which is the opposite of the emergency-triage framing usually used in vendor communications. Reported algorithm performance: sensitivity 88.4%, specificity 96.1%, PPV 85.9%.

Technology
AI technique
Deep learning, Computer vision
Input data
Medical imaging
Output type
prioritised_list
Autonomy level
Informs a human (advisory)
Model provenance
Vendor proprietary
Deployment
Country
United States
Regulatory & governance
Data identifiability
Not stated in the source
Performance summary
Headline metric
Other
Value
90

Similar deployments

AI
Looking for close matches in the Library…

Evidence records

Studies and evaluations attached to this use case.

  • implementation_study

    Reduction in scan view delay, outpatients (%): 90

    Population: 8,723 head CTs, single academic centre

    Source

  • implementation_study

    Reduction in scan view delay, emergency setting (%) — not statistically significant (p=0.37): 15

    Population: 8,723 head CTs, single academic centre

    Source

  • implementation_study

    Sensitivity for intracranial haemorrhage (%): 88.4

    Population: 8,723 head CTs, single academic centre

    Source

  • implementation_study

    Specificity for intracranial haemorrhage (%): 96.1

    Population: 8,723 head CTs, single academic centre

    Source

Study details

Study type
Retrospective single-centre cohort
Sample size
8723 participants

Contributors

Deploying organisation
University of Chicago Medical Center · Hospital / health system · United States
AI vendor
Aidoc Medical, Ltd. · Israel
Product name
BriefCase
Source record
PubMed · PMID:34201775