Table 4.

Model performance of deep learning algorithms.

Clinical outcome	Imaging modality	Clinical variable	Model	References	Sample size (T/EV)	Model performance		Validation
						AUC (95% CI)	Others	Internal	External
Good functional outcome at 90 days (mRS ≤ 2)	CTA	No	DL (RFNN)	Hilbert et al. (34)	1,301	0.71(0.68–0.74)†	n.a.	Yes	No
	NCCT	Yes	DL (CNN)	Samak et al. (35)	400 (hold-out testing: 100)	0.75 (0.63–0.87)†	ACC,0.77	Yes	No
	MRI (DWI)	No	DL (CNN)	Nishi et al. (36)	250/74	Internal: 0.81 (0.70–0.92)† External:0.73 (0.61–0.85)†	Internal: SEN,0.76; SPE,0.76; ACC,0.72; External: SEN,0.72; SPE,0.60; ACC,0.65	Yes	Yes
Multiclass mRS (0, 1, 2, 3, 4, 5, 6) at 90 days	NCCT	Yes	DL (CNN)	Samak et al. (35)	400 (hold-out testing: 100)	n.a.	ACC, 0.35	Yes	No
Successful reperfusion (TICI score≥2b)	CTA	No	DL (RFNN)	Hilbert et al. (34)	1,301	0.65 (0.62–0.68)†	n.a.	Yes	No
Haemorrhagic transformation (including HI1, HI2, PH1, and PH2)	MRI (DWI and PWI)	Yes	DL (CNN)	Jiang et al. (37)	338/54	Internal: 0.95 (0.87–1.00)† External:0.94 (0.85–1.00)†	Internal: SEN, 0.86; SPE, 0.90; ACC,0.89; External: SEN,0.86; SPE,0.89; ACC,0.88	Yes	Yes

ANN, artificial neural networks; AUC, area under the Receiver Operating Characteristic curve; ACC, accuracy; CTA, computed tomography angiography; CNN, convolutional neural network; DWI, diffusion weighted imaging; EV, external validation dataset; HI, hemorrhagic infarction; PH, parenchymatous hematoma; MRI, magnetic resonance imaging; NCCT, non-contrast computed tomography; NV-, negative predictive value; PWI, perfusion weighted imaging; PV+, positive predictive value; RFA, random forest analysis; RLR, regularized logistic regression; RFNN, receptive field neural networks; SVM, support vector machine; SEN, sensitivity; SPE, specificity; T, training dataset; n.a., not available/not applicable. Note: *model derived from patients registered in MR CLEAN Registry (38).^†95% CI was estimated based on normal distribution.