Design and Development of Serial Ensemble Deep Network and Enhanced Fusion Strategies for Parkinson’s Disease Analysis

Neurodegenerative disease is a progressive disorder characterized by degeneration of nerve cells in the brain and spinal cord, that producing cognitive, behavioral, and motor impairments.However, primary and exact diagnosis is a crucial process for effective handling, therapy, Deep Brain Stimulation (DBS) and medication to enhance quality of human daily life.In Parkinson's Disease (PD) analysis, leveraging different data modalities are used like neuroimaging, voice data, clinical features, and image samples.Deep learning (DL) presents significant performance in exact diagnosis and classification, but it faces certain limitations due to uneven data samples and high dimensional limits clinical utilization.This research study proposes a novel serial ensemble deep network and enhance fusion mechanism for predicting multi-depression score by multimodal data samples.Initially, genetic biomarkers and input images are processed with Z-score normalization and Improved Homomorphic Gaussian Filter (IHGF) to enhance quality and remove the noise background, which can provide effective data samples.Tabular based Additive Attention assisted 1DConvNeXT (TA 2 ConvNeXT) Network to capture significant features from genetic biomarkers.Then, preprocessed image is processed with Densely Bidirectional Gated based VGG Network (DBiG_VGGNet) and Improved Global Attention based Residual-Transformer Network (IGA_ResTNet) for extracting relevant features and decrease dimensionality problem.Finally, Cross Attention based Learnable Feature Fusion (CAF 3 ) module to fuse genomic and image features for predicting multi-output regression GDS and UPDRS-I scores.Experimental study results show model achieves Mean Absolute Error (MAE) value of 0.089 and Mean Square Error (MSE) value of 0.057 when compared to existing models, proving its effectiveness in multiple PD depression score prediction.