Ischemic stroke studies using artificial intelligence (AI) for acute diagnosis, triage, or complication prediction with performance compared with humans (2014–2019). Controls for bias and possible sources of bias for each study are shown in the online supplementary table I
| Study | Aim | Inclusion and exclusion criteria | Study type | AI method and Software | Secondary validation, | Accuracy metrics | n | Conclusions |
| Maegerlein et al 201924 | AI vs neuroradiologist ASPECTS | Acute MCA with or without an LVO | SCRV | RFL, RAPID | Two neuroradiologists | Kappa human consensus: ~57%, AI 90% | 152 | ASPECTS by AI has better agreement to consensus score than independent neuroradiologists |
| Chatterjee et al 201929 | CNN to detect LVOs from CTA images | Anterior circulation stroke syndromes | SCRV | CNN, Viz.ai | Radiology report without AI | Sensitivity 82%, specificity 94%, PPV 77%, NPV 95% | 650 | CNNs detect LVOs with good accuracy |
| Öman et al 201914 | CNN to detect MCA strokes from CTA source images | Stroke code patients: MCA cases and final diagnosis of non-stroke controls | SCRV; two-part AI development and validation | CNN, 3D slicer, DeepMedic | Two radiologists’ manual segmentation (ASPECTS) and quantitative28 voxel score | Sensitivity 67–74%, specificity 93–96%, AUC 91–93% | 60 | CNNs can use CTA source images to detect hypodensities correlated to acute ischemic stroke |
| Kuang et al 201917 | ASPECTS vs MRI DWI | Acute ischemic stroke CT head and then MRI DWI brain | SCRV | RFL | MRI DWI sequences | Kappa 88%, sensitivity 98%, specificity 80% | 157 | RFL reliably assigns ASPECTS |
| Seker et al 201925 | AI vs neuroradiologist ASPECTS with different CT postprocessing reconstruction | Acute proximal MCA occlusion | SCRV | RFL, Brainomix e-ASPECTS | Four radiologists | e-ASPECTS performed better than residents: kappa 92% vs ~75% | 43 | e-ASPECTS performs better than radiologists with varied reconstructions of CT scans from the same patient |
| Olive-Gadea et al 201926 | AI vs radiologist ASPECTS association with CTP infarct core | Acute proximal MCA or ICA terminus occlusions that received ET | SCRV | RFL, Brainomix e-ASPECTS, RAPID CTP | RAPID CTP core | Radiologist vs AI ASPECTS correlation 44%, both scores correlated with CBV (rs = –0.41 vs –0.54) | 184 | Radiologist and AI ASPECTS compared with CBV value similarly predict infarct volume |
| Chriashkova et al 201927 | AI vs radiologist ASPECTS and time to score | Acute ischemic stroke syndromes | SCRV | RFL, Brainomix e-ASPECTS | 26 Clinicians (including 11 radiologists) with access to clinical symptoms and follow-up scans | AI reduced scoring time by 34%, kappa improved from 26% to 38% | 64 | AI in e-ASPECTS improved efficiency and accuracy |
| Barreira et al 201813 | CNN vs neurologist to detect LVOs from CTA | Acute ischemic stroke with or without LVO | Multicenter retrospective study | CNN, Viz.ai | One stroke neurologist | Sensitivity of 97% and specificity of 52%, PPV 74%, NPV 91%; overall accuracy 78% | 152 | CNN have high positive predictive value for presence of an LVO |
| Barreira et al 201812 | CNN vs neurologists to detect LVOs from CTA | Acute ischemic stroke with or without LVO | Three-center retrospective review | CNN, Viz.ai | Several experienced stroke neurologists | Sensitivity 90%, specificity 83%, accuracy 86%, PPV 82%, NPV 91%, AUC 86% | 875 | CNN AI detects more than 82% of anterior circulation LVOs |
| Boldsen et al 201816 | AI vs radiologist for an ATLAS core stroke | Acute anterior circulation stroke, MRI DWI scan | Multicenter, retrospective case training, validation | RFL | One imaging expert | Dice coefficient 0.61 | 109 | AI provides reliable penumbra and core size |
| Kellner et al 201832 | AI vs clinical scales to identify severe stroke types, including LVO | Any brain neurological presentation with neurological imaging | Multicenter, combined pilot and validation study | ML, subtype not defined | Multiple trial coordinators and clinicians | Sensitivity 93%, specificity 87%, AUC 95% | 252 | Volumetric impedance phase shift spectroscopy run by ML reliably identifies severe strokes |
| Lucas et al 201819 | CNN vs radiologist to predict CTP stroke growth | Patients with a stroke, each with a full set of CTP data | SCRV AI development, validation | CNN | One expert manual rater | Dice coefficient 0.46–0.53 | 29 | CNN predicts infarct growth well |
| Goebel et al 201822 | ASPECTS from: AI softwares vs each other vs radiologist | Acute MCA ischemia, no hemorrhage or artifacts | SCRV, head-to-head scoring | Frontier vs Brainomix vs radiologist | Two senior radiologists | Agreement interclass correlation: human and Brainomix 0.71–0.84 vs human and frontier 0.47–0.54 | 150 | High ASPECTS agreement between consensus radiologists and Brainomix |
| Guberina et al 201823 | AI vs radiologist ASPECTS with acute and chronic strokes | Acute MCA ischemia within 6 hours from symptom onset who were eligible for reperfusion therapy | SCRV; definite infarct core on head CT at 24 hours | RFL, Brainomix e-ASPECTS | Three neuroradiologists | AI sensitivity 83% vs human 73%, specificity 57% vs 84% | 119 | RFLhas slightly higher sensitivity but worse specificity for ASPECTS |
| Nagel et al 201710 | AI vs radiologist ASPECTS | Acute anterior circulation ischemic stroke syndrome, baseline and 24–36-hour follow-up head CT scans | Retrospective, multicenter comparison study | RFL, Brainomix e-ASPECTS | Three radiologists | AI sensitivity ~45%, specificity ~92%, accuracy ~86% | 132 | RFL was non-inferior to neuroradiologists for ASPECTS |
| Lisowska et al 201728 | CNN vs radiologist acute stroke hypodensity and dense vessel sign detection | Acute ischemic stroke within 6 hours from symptom onset with non-contrast CT | Retrospective, pooled patients from three studies | CNN, Keras on Theano (Python) | One researcher, one radiologist | AI AUC 92–96% | 51 | CNN detection of asymmetrical hypodensity and dense vessel can detect ischemia well |
| Herweh et al 20169 | AI vs neurologist ASPECTS | Acute stroke, CT and MRI with DWI images <2 hours apart | Retrospective single center study | RFL, Brainomix e-ASPECTS | Three stroke experts and three neurology trainees, DWI as gold standard | AI sensitivity 46%, specificity 95%. AI correlation coefficient 0.44 vs humans 0.19–0.38 | 34 | RFL accuracy for ASPECTS was similar to neurologists |
| Bentley et al 201437 | AI vs radiologist prediction of sICH, mRS | Acute ischemic stroke treated with IV tPA | SCRV, moderate-supervision of AI | SVM | Radiologist SEDAN and HAT scoring | Radiologist 67% vs AI 74% AUC | 112 | AI is superior to radiologist scales for sICH |
| Takahashi et al 201431 | AI vs neuroradiologist detection of MCA dot sign | Anterior ischemic stroke, first head CT within 24 hours | Two-center, retrospective study | SVM | Two neuroradiologists | AI sensitivity 98%, false positive rate 1.3% per image | 7 | AI SVM is accurate to detect the MCA dot sign |
| Thornhill et al 201433 | AI vs neuroradiologist intraluminal thrombus from plaque | Patients with acute TIA or stroke with suspected free-floating luminal thrombus on CTA | SCRV, training and ML validation | ANN, SVM | Two neuroradiologists | SVM sensitivity 88%, specificity 71%, AUC 85%. | 23 | AI is ableto distinguish shape signatures of intraluminal thrombus from plaques |
AI, artificial intelligence; ANN, artificial neural network; ASPECTS, Alberta Stroke Programme Early CT Score; AUC, area under the curve; CBV, cerebral blood volume; CNN, convolutional neural network; CTA, CT angiography; CTP, CT perfusion; ET, endovascular thrombectomy; ICA, internal carotid artery; LINDA, lesion identification with neighborhood data analysis; LVO, large vessel occlusion; MCA, middle cerebral artery; ML, machine learning; mRS, modified Rankin Scale; NPV, negative predictive value; PPV, positive predictive value; RFL, random forest learning; SCRV, single-center, retrospective review; sICH, symptomatic intracerebral hemorrhage; SVM, support vector machine (form of artificial network); TIA, transient ischemic attack; tPA, tissue plasminogen activator.