Fusion based AI Model for Spinal Cord Outcome Prediction in Diabetic Patients

Magnetic Resonance Imaging (MRI) of the spine, when integrated with diabetesrelated clinical data offers significant for improving disease prediction and recovery. This study develops a multimodel deep learning framework that combines spinal MRI images with diabetes-related tabular data such as glycated hemoglobin (HbA1c), glucose level, and body mass index (BMI) to assess both visual-based and clinical indicators of patient health. A ResNet50-based Convolutional Neural Network (CNN) is utilized to extract deep visual features from spinal images; whereas, diabetes-related tabular data will be processed through a neural network that is fully connected. Heterogeneous features are merged in concatenation layers, which are optimized jointly in a single model using the Adam optimizer with categorical crossentropy loss. The model is trained and fine-tuned to learn the complex relationships between changes in spinal structure and conditions related to diabetes, which produces richer diagnostic insights. The experimental results show that this multimodal approach performs better than single modality. It provides better accuracy and generalization when predicting disease severity and recovery outcomes in post-operative patients. The multimodal deep learning framework also has tools for interpretability, like Grad-CAM and SHAP, to visualize feature importance from images and tabular data. These results indicate that the developed framework emphasizes the unique potential of using multiple types of medical data. This can improve personalized and data-driven decision making for clinical diagnosis and prognosis prediction.