Neurodegenerative Condition Detection Using Modified Metaheuristic for Attention Based Recurrent Neural Networks and Extreme Gradient Boosting Tuning

Abstract

Parkinson’s disease is a neurological disorder, caused by the death of dopaminergic neurons which can cause various movement disorders to appear, recognized as standard Parkinson’s motor symptoms. A drug to stop the progression of the disease is very difficult to find, so current treatment is based on alleviating the symptoms of the disease itself. As no direct treatment exists that would cure the condition, early detection and proper treatment are essential in maintaining the patient’s quality of life. This work explores the potential of merging artificial intelligence and machine learning algorithms for Parkinson’s disease early detection from finger-tapping accelerometer tests. Time series classification is explored through the use of recurrent neural networks augmented with and without attention layers. Additionally, extreme gradient boosting in combination with statistical analysis is explored in order to differentiate Parkinson’s from other developing neurodegenerative disorders. As the performance of algorithms hinges on proper parameter selection, this work applies metaheuristics for performance optimization. A modified version of a recently introduced sinh cosh optimizer algorithm is also proposed. The approach is tested on a publicly available real-world clinical dataset consisting of patients and control group samples and a total of three separate experiments were conducted. The introduced optimizer demonstrated admirable performance in comparative analysis, with the best performing models exceeding 90% accuracy.

Keywords

• Parkinson’s disease
• metaheuristics
• sinh cosh optimization
• optimization
• hyperparamater tuning