TABLE 2.

Performance of deep learning algorithm for predicting motor outcome after corona radiata infarct.

	MBC prediction model	FAC prediction model
Sample size (patients)	154 For training (462 images, 70%), 66 for validation (198 images, 30%) Sample ratio: Poor 52.7% (0), good 47.3% (1)	154 For training (462 images, 70%), 67 for validation (201 images, 30%) Sample ratio: Poor 54.3% (0), good 45.7% (1)
CNN model	Model for MR images
	- EfficientNetB0 with fine-tuning - SGD optimizer, ReLU activation - Data augmentation and dropout for regularization - Image of size 256 × 256	- EfficientNetB0 with fine-tuning - RMSProp optimizer, ReLU activation - Data augmentation and dropout for regularization - Image of size 256 × 256
Sequential neural network model	Model for clinical data - 3 hidden layers with 256-512-1,024 neurons - SGD optimizer, ReLU activation - Batch normalization for regularization - 11 clinical variables as inputs	- 3 hidden layers with 256-512-1,024 neurons - RMSProp optimizer, ReLU activation - Batch normalization for regularization - 11 clinical variables as inputs
Integrated prediction model	Concatenated model with CNN and sequential neural network outputs - MBC and FAC prediction with three images and clinical data per patient - Binary classification with sigmoid activation
Decision criteria for integrated prediction model	Poor (0): less than 3 “good” predictions; good (1): 3 “good” predictions
Integrated prediction model performance	MBC prediction accuracy of 90.91% on training data Training AUC of 0.907 with 95% CI [0.861–0.953] MBC prediction accuracy of 89.39% on validation data Validation AUC of 0.891 with 95% CI [0.814–0.967]	FAC prediction accuracy of 91.6% on training data Training AUC of 0.935 with 95% CI [0.896–0.975] FAC prediction accuracy of 91.1% on validation data Validation AUC of 0.919 with 95% CI [0.842–0.995]

MBC, modified Brunnstrom classification; FAC, functional ambulation category; MR, magnetic resonance; CNN, convolutional neural network; SNN, sequential neural network; SGD, stochastic gradient descent; ReLU, rectified linear unit; RMSProp, root mean square propagation; AUC, area under the curve; CI, confidence interval.