TABLE 4.

Ablation study of the integrated prediction model.

	MBC prediction model	FAC prediction model
Integrated model	Integrated model with both MR images and clinical data as input EfficientNetB0 for MR images Sequential neural network with three layers (1024-512-256) for clinical data Accuracy: 90.91% on training data, 89.39% on validation data AUC: 0.907 on training data, 0.891 on validation data	Integrated model with both MR images and clinical data as input EfficientNetB0 for MR images Sequential neural network with three layers (1024-512-256) for clinical data Accuracy: 91.6% on training data, 91.1% on validation data AUC: 0.935 on training data, 0.919 on validation data
CNN model only	EfficientNetB0 CNN model with fine tuning MR images as input Training parameters: SGD optimizer, ReLU activation, lr 8e-06, dr 0.2, bs 64, binary classification with sigmoid activation Accuracy: 73.8% on training data, 63.6% on validation data AUC: 0.974 on training data, 0.619 on validation data	EfficientNetB0 CNN model with fine tuning MR images as input Training parameters: RMSProp optimizer, ReLU activation, lr 8e-06, dr 0.25, bs 64, binary classification with sigmoid activation Accuracy: 72.2% on training data, 63.0% on validation data AUC: 0.852 on training data, 0.662 on validation data:
SNN model only	SNN with clinical data 3 Hidden layers with 256-512-1024 neurons Training parameters: SGD optimizer, ReLU activation, lr 8e-06, dr 0.2, bs 64, binary classification with sigmoid activation Batch normalization and dropout for regularization 11 Clinical variables as inputs Accuracy: 94.2% on training data, 83.3% on validation data AUC: 0.980 on training data, 0.845 on validation data	SNN with clinical data 3 Hidden layers with 256-512-1024 neurons Training parameters: RMSProp optimizer, ReLU activation, lr 8e-06, dr 0.25, bs 64, binary classification with sigmoid activation Batch normalization and dropout for regularization 11 Clinical variables as inputs Accuracy: 96.3% on training data, 80.6% on validation data AUC: 0.992 on training data, 0.785 on validation data:

ML, machine learning; 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; lr, learning rate; dr, dropout rate; bs, batch size.