Early diagnosis and developmental outcome prediction of agenesis of the corpus callosum via an interpretable deep multimodal fusion model

Objective Agenesis of the corpus callosum (ACC) presents with highly heterogeneous clinical features. Common methods rarely achieve accurate prenatal or early postnatal diagnosis and prognosis. We aimed to develop and test an interpretable deep neural network (DNN) that combines multimodal clinical data to improve diagnostic accuracy and neurodevelopmental outcome prediction. Methods We collected data from 205 pediatric patients with ACC at Wuhan Children’s Hospital between 2016 and 2024. A total of 27 clinical features were extracted, including neuroimaging findings, perinatal risk factors, and follow-up developmental quotients (Gesell Developmental Schedules and Gross Motor Function scores). Five-fold cross-validation was adopted. We built an eight-layer fully connected DNN with ReLU activation in the hidden layers. For categorical endpoints, a sigmoid output layer with binary cross-entropy loss was used. For continuous endpoints, a linear output layer with mean squared error loss was used. SHAP (Shapley Additive Explanations) values were used to quantify the contribution of individual features to model predictions. Performance was compared with a support vector machine (SVM) baseline and across hyperparameter settings. Area under the receiver-operating-characteristic curve (AUC), F1 score, precision, recall, mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R 2 ) served as primary metrics. Results Across 12 neurodevelopmental disorders, the model reached an average AUC of 0.97. AUCs for intellectual disability, autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD), specific learning disorder and developmental coordination disorder ranged from 0.98 to 1.00. Prediction remained moderate for cerebral palsy (AUC = 0.74) and epilepsy (AUC = 0.67). MAE for both Gesell and Gross Motor Function scores was 0.10, with corresponding R 2 values of 0.62 and 0.63. SHAP analysis identified extracranial malformation (clinical type III), facial dysmorphism and birth weight as the most influential features for developmental outcome. The DNN model outperformed the SVM baseline, with an AUC improvement of 0.16 for communication disorder and an R 2 increase of 0.19 for Gesell score ( p 0.001). Ablation experiments confirmed eight layers, sixteen neurons per layer, a learning rate of 0.01 and ten training epochs as the optimal configuration. Additional layers or higher learning rates caused overfitting. Conclusion The proposed interpretable DNN framework outperforms traditional classifiers in early ACC diagnosis and developmental outcome prediction. It provides a potential tool for clinical decision support. Larger samples and integration of raw imaging data are needed to enhance prediction of complex phenotypes such as cerebral palsy and epilepsy.