Table 1

Clinical AI reports

Name of device or algorithm
Brief descriptionData collection methodsFDA approval statusType of algorithmData set compositionPopulation ethnic compositionBias assessment evaluationModel evaluation/Research protocolMetrics for performance errors* †Clinical workflow implementation
RadiologyIntelDecision support software to augment medical imaging-related diagnosisStandard H&E stained images, stimulated Raman histology510(k) Premarket notificationConvolutional neural networkSize/Composition of training dataset: 550 000 inpatients, academic medical centres
Size/Composition of testing dataset: 350 000
inpatients at community hospitals
Non-Hispanic white 60%
Hispanic and Latino 18%
Black/African-American 13%
Asian 6%
Other 3%
Google TCAV
Multi-centred prospective clinical trial and retrospective analysisArea under the curve 0.85
Classification accuracy 75%
Integrated into 50 hospitals via EHR systems, including Epic, Cerner
DiabetEYECDS system to enhance screening/diagnosis of diabetic retinopathyWidefield stereoscopic photography and macular optical coherence tomographyDe novo pathwayConvolutional neural networkSize/Composition of training dataset: 7000 outpatients, primary care clinic
Size/Composition of testing dataset: 5000
Outpatients at independent clinic
Non-Hispanic white 70%
Hispanic and Latino 10%
Black/African-American 10%
Other 10%
None availableRandomised controlled trialSensitivity, 81%, specificity, 90%,
Area under the curve 0.80
Confusion matrix 0.91
Implemented in 150 primary care clinics in the USA
  • *Mishra.19

  • †Scott et al.20

  • AI, artificial intelligence.