J Med Life Sci > Volume 20(4); 2023 > Article |
|
Study | Subjects | ML model | Image modality | Ground-truth | Result |
---|---|---|---|---|---|
Kuang at al. (2019) [69] | 157 AIS <8 hours | Random forest | NCCT* | DWI ASPECTS | AUC, 0.79; κ=0.6; sensitivity, 66.2%; specificity, 91.8% |
AUC, 0.89; κ=0.78; sensitivity, 97.8%; specificity, 80% (for dichotomized ASPECTS <4 and ≤4) | |||||
Qiu at al. (2020) [77] | 157 AIS <6 hours | U-Net | NCCT* | DWI lesion volume | r=0.76, P<0.001 |
Random forest | Mean difference, 11 mL | ||||
McNemar P=0.11 for prediction of volume >70 mL | |||||
Tuladhar et al. (2020) [118] | 204 AIS | 3D multi-scale CNN | NCCT* | NCCT manual segmentation | DSC, 0.47 (DSC, 0.5 after post-processing) |
Lesion volume ICC=0.88 | |||||
Chen at al. (2022) [75] | 1,391 AIS <6 hours | DCNN | NCCT* | NCCT manual segmentation | DSC, 0.6; AUC, 0.876 in internal validation set |
DSC, 0.762; AUC, 0.729 in external validation set | |||||
Winzeck et al. (2019) [10] | 116 AIS <24 hours | Ensemble of CNN | DWI, ADC, LOWB† | Manual segmentation | Volume, r=0.91 |
Outline segmentation, κ=0.86-0.90 | |||||
Kim at al. (2019) [79] | 296 AIS within 3 days | U-Net | DWI, ADC† | Manual segmentation in DWI | ICC with manual segmentation, 1.0 (95% CI, 0.99-1.00) |
ICC with RAPID, 0.99 (95% CI, 0.98-0.99) | |||||
Woo et al. (2019) [80] | 246 AIS | 2D U-Net, DenseNet | DWI† | Manual segmentation in DWI | DSC, 0.85 for U-Net and DenseNet |
DSC, 0.86 for ensemble of U-Net and DenseNet | |||||
DSC, for small lesions 0.82 | |||||
DSC, for large lesions 0.89 |
CT: computed tomography, MRI: magnetic resonance imaging, ML: machine learning, AIS: acute ischemic stroke, NCCT: non-contrast computed tomography, DWI: diffusion-weighted image, ASPECTS: Alberta stroke program early CT score, AUC: area under the curve, CNN: convolutional neural networks, DSC: dice score coefficient, ICC: intraclass correlation coefficient, DCNN: deep convolutional neural network, ADC: apparent diffusion coefficient, LOWB: low b-value diffusion-weighted image, CI: confidence interval.
Study | Number of patients | ML model | Imaging information | Clinical information | Outcome | Results |
---|---|---|---|---|---|---|
Heo et al. (2019) [15]* | 2,604 | DNN | None | Demographics | Favorable outcome (3mo mRS, ≤2) | DNN AUC, 0.888 |
Random forest | NIHSS | Random forest AUC, 0.857 | ||||
LR | Stroke subtype | LR AUC, 0.849 | ||||
Medical history | ASTRAL AUC, 0.839 | |||||
Time from onset to admission | ||||||
Laboratory | ||||||
Sung et al. (2021) [114]* | 3,847 | DL using BERT | CT reports | HPI | Poor outcome (3mo mRS, >2) | HPI+CT AUC, 0.840 |
NIHSS AUC, 0.811 | ||||||
PLAN AUC, 0.837 | ||||||
ASTRAL AUC, 0.840 | ||||||
Brugnara et al. (2020) [119]* | 246 | Gradient boosting classifiers | Lesion volume | Demographics | Favorable outcome (3mo mRS, ≤2) | Clin+imaging+Angio+postintervention charact AUC, 0.856 |
eASPECTS | Medical history | |||||
Vessel occlusion site | Laboratory | Clin+imaging AUC, 0.740 | ||||
e-CTA collateral score | NIHSS | |||||
Perfusion parameters | Premorbid mRS | |||||
mTICI | Angiographic time parameter | |||||
Quan et al. (2021) [120]* | 110 | PyRadiomics for feature selection | FLAIR | Demographics | Poor outcome (3mo mRS, >2) | Rad+Clin+Con MR AUC, 0.926 |
ADC radiomics vs. lesion diameter | Medical history | Rad AUC, 0.815 | ||||
LR | DWI-ASPECTS | NIHSS | Clin AUC, 0.791 | |||
Fazekas score | Time from onset to image | Clin+Con MRI AUC, 0.782 | ||||
Kim et al. (2021) [110]* | 2,363 | SVM | Lateralization | Demographics | END | LightGBM AUC, 0.772 |
XGBoost LightGBM | Lesion pattern | Medical history | LR AUC, 0.696 | |||
SVS | Laboratory | |||||
MLP | Atherosclerosis | NIHSS | ||||
LR | HT | |||||
Moulton et al. (2023) [121]* | 322 | 3D CNN | DWI | None | 3mo mRS, >2 | DL AUC, 0.83 |
LR | Lesion volume AUC, 0.78 | |||||
ASPECTS AUC, 0.77 | ||||||
Meinel et al. (2022) [112]* | 2,261 | XGBoost | White matter disease | Demographics | Futile recanalization (3mo mRS, 5-6) | XGBoost AUC, 0.87 |
Medical history | ||||||
Vessel occlusion | Laboratory | |||||
Treatment | ||||||
Zeng at al. (2022) [113]* | 110 | SVM | CT lesion volume | Demographics | Futile recanalization (3mo mRS, 3-6) | Futile recanalization AUC, 0.945 |
RFC | Hypo- and hyperdense lesions levels | Medical history | Cerebral edema AUC, 0.885 | |||
XGBoost | NIHSS | Cerebral edema | Cerebral herniation AUC, 0.904 | |||
KNN | GCS | Cerebral herniation | ||||
GBM with LR-Stacking | Laboratory | |||||
Interventional time parameters | ||||||
Treatment | ||||||
Xu et al. (2019) [111]† | 5,159 | XGBoost LR | None | Demographics | IS or TIA readmission within 90 days | XGBoost AUC, 0.782 |
Medical history | LR AUC, 0.771 | |||||
Length of stay | ||||||
Laboratory | ||||||
Ntaios et al. (2021) [122]† | 2,832 | XGBoost | None | Demographics | Major adverse cardiovascular event | XGBoost AUC, 0.648 |
Random forest SVM | Medical history | |||||
Medication | ||||||
Wang et al. (2022) [123]† | 1,003 | Recurrent neural network | DWI | Demographics | Recurrent IS at 1 year | Clin+Rad AUC, 0.847 |
ADC radiomics | Medical history | Clin AUC, 0.675 | ||||
NIHSS | Rad AUC, 0.779 | |||||
TOAST |
ML: machine learning, DNN: deep neural network, LR: logistic regression, NIHSS: National Institutes of Health stroke scale, 3mo mRS: 3-month modified Rankin scale, AUC: area under the receiver operating characteristic curve, ASTRAL: acute stroke registry and analysis of Lausanne, DL: deep learning, BERT: bidirectional encoder representations from transformers, CT: computed tomography, HPI: history of present illness, PLAN: preadmission comorbidities, level of consciousness, age, and neurological deficit, ASPECTS: Alberta stroke program early CT score, mTICI: modified thrombolysis in cerebral infarction, Clin: clinical information, Angio: angiography, FLAIR: fluid-attenuated inversion recovery, ADC: apparent diffusion coefficient, DWI: diffusion weighted image, Rad: radiomics, Con MR: conventional MRI features, MRI: magnetic resonance imaging, SVM: support vector machine, XGBoost: extreme gradient boosting, LightGBM: light gradient boosting machine, MLP: multilayer perceptron, SVS: susceptibility vessel sign, HT: hemorrhagic transformation, END: early neurological deterioration, CNN: convolutional neural networks, RFC: random forest classifier, KNN: K-nearest neighbor, GBM: gradient-boosting machine, GCS: Glasgow coma scale, IS: ischemic stroke, TIA: transient ischemic attack, TOAST: trial of Org10172 in acute stroke.
Mi-Yeon Eun
https://orcid.org/0000-0002-8617-5850
Eun-Tae Jeon
https://orcid.org/0000-0002-7000-9719
Jin-Man Jung
https://orcid.org/0000-0003-0557-6431