This study presents an automated Parkinson's disease severity prediction framework that combines ensemble feature extraction (IAOA-weighted features, RBM, t-SNE) with an Adaptive Multi-scale Temporal Convolutional Network (AMTCN) to classify UPDRS-based severity, reporting ~94% accuracy.

Improved and more interpretable severity prediction can aid clinical monitoring, patient stratification, and trial enrollment, but the work offers little direct mechanistic or therapeutic insight for drug discovery.

RESEARCH BACKGROUND: Parkinson's Disease (PD) requires accurate severity prediction models for enabling efficient treatment planning and disease management. The recent improvements in deep learning have shown promising results while predicting PD severity but, conventional models encounter various complexities in analyzing the informative features, managing complex data patterns and tuning the model parameters. PROBLEM FORMULATION: The problems in the prior research studies are addressed by implementing an effective and robust deep learning approach that can accurately capture and identify the complicated relations and patterns in the PD data. MAJOR CONTRIBUTION: The primary contribution of the research study is implemented by employing an adaptive multi-scale deep learning model that combines the ensemble feature extraction and optimization model. METHODOLOGY: An automated PD severity prediction framework is introduced in this research study. Initially, the required amount of data are collected and given into the ensemble feature extraction process. During this process, the three feature sets, such as optimal weighted features, Restricted Boltzmann Machine (RBM) features, and "t-distributed Stochastic Neighbor Embedding (t-SNE) features" are extracted, whereas the optimal weighted features are obtained by the Improved Archimedes Optimization Algorithm (IAOA). Here, the newly improved IAOA by a new random number-based strategy, which improves the convergence rates and supports the parameter tuning process. The framework addresses clinical interpretability by leveraging the optimal weighted features generated by the IAOA. This feature selection process ensures that the most relevant features are identifiable to clinicians, building clinical trust and supporting decision-making. Due to this transparency, the developed framework moves beyond the typical black-box nature of deep learning models. Subsequently, the Adaptive Multi-scale Temporal Convolutional Network (AMTCN) model is designed to predict the PD severity. Here, this network is used to predict an individual's PD severity by employing diverse ranges of the Unified Parkinson's Disease Rating Scale (UPDRS) as a class label. Moreover, the IAOA algorithm is used for tuning the AMTCN parameters. Finally, the distinct experimental outcome is conducted by contrasting with various conventional approaches. RESULTS: The outcome of the developed model shows 94% in terms of accuracy and specificity higher than the conventional models. Therefore, this improved performance helps in reducing the diagnostic error rate, resulting in more precise and comprehensive evaluation.