Table 2 Sensitvity and specificity for individual particpants, swarm sessions, and AI models.
From: Human–machine partnership with artificial intelligence for chest radiograph diagnosis
Participants | Diagnostic performance parametersa | |
|---|---|---|
Sensitivity | Specificity | |
Swarm sessions | ||
Group A (N = 7) | ||
Individual average | 0.642 [0.579, 0.709] | 0.819* [0.777, 0.862] |
Crowd-based majority | 0.650 [0.412, 0.783] | 0.900 [0.800, 0.972] |
Crowd-based mean probability | 0.650 [0.462, 0.824] | 0.900 [0.806, 1.00] |
Swarm interpolation | 0.700 [0.526, 0.875] | 0.933 [0.852, 1.00] |
Group B (N = 6) | ||
Individual average | 0.633 [0.558, 0.704] | 0.883 [0.845, 0.920] |
Crowd-based majority | 0.600 [0.421, 0.789] | 0.933 [0.846, 1.00] |
Crowd-based mean probability | 0.600 [0.421, 0.789] | 0.933 [0.846, 1.00] |
Swarm interpolation | 0.700 [0.500, 0.867] | 0.933 [0.844, 1.00] |
Combined (N = 13) | ||
Individual average | 0.519ϕ† [0.471, 0.568] | 0.690 [0.654, 0.724] |
Crowd-based majority | 0.625ϕØ [0.477, 0.721] | 0.917 [0.857, 0.968] |
Crowd-based mean probability | 0.625ϕ† [0.500, 0.744] | 0.917Ø [0.852, 0.968] |
Swarm interpolation | 0.700Ψ [0.578, 0.814] | 0.933ϕ [0.855, 0.968] |
Deep learning models | ||
CheXNet | 0.450*‡† [0.326, 0.579] | 0.867 [0.793, 0.932] |
CheXMax | 0.900^‡† [0.773, 1.00] | 0.767*‡† [0.672, 0.857] |
Augmented HITL model (combined swarm and CheXMax) | 0.875*‡† [0.783, 0.956] | 0.933ϕ [0.877, 0.983] |
