Date of Award

7-8-2024

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.K.Ramkumar

Keywords

Explainable Artificial Intelligence, Machine Learning, Gait Signals, Parkinson’s Disease, Deep Learning

Abstract

Parkinson’s disease (PD) is a degenerative neurological condition marked by motor symptoms like tremors, bradykinesia, and stiffness. It is observed that early and precise diagnosis of this disease is crucial, as that will have a significant impact on effective disease management and intervention. This thesis explores the technical feasibility of applying AI techniques to recognize patterns from spiral and wave drawings, which are usually a unique signature type for PD patients.

The focus of the work is to diagnose the disease through novel deep transfer learning techniques, to diagnose the severity of the disease, and also to develop effective model interpretability techniques to improve the reliability of the trained models. The study gathers spiral and wave drawing data from PD patients and healthy controls from standard benchmarked data sets, followed by applying standard preprocessing and training techniques to extract relevant characteristic features.

Advanced transfer learning techniques, such as fine-tuning and pre-trained convolutional neural network (CNN) models, are applied to identify PD-related patterns automatically. Model interpretability techniques are then employed to understand learned representations and discover classification aspects, enhancing the identification of significant motor and non-motor symptoms linked to PD-related drawing anomalies. The proposed approach showcases a comprehensive hybrid deep transfer learning model incorporating explainable AI (XAI), which has been created for the early detection of Parkinson’s Disease (PD), achieving a classification accuracy of 98.45%.

Furthermore, this thesis evaluates the effectiveness of gait signals in PD diagnosis using recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU) architectures. Data collected from PD patients and healthy individuals are pre-processed, and relevant features are extracted for model training. The models are trained to identify PD-related patterns autonomously. Each model is evaluated based on accuracy, sensitivity, and specificity. Results indicate that RNN, LSTM, and GRU architectures effectively diagnose PD using gait signals, accurately differentiating between PD and non-PD instances. The GRU model achieves a higher classification accuracy of 98.20% in accurately classifying PD. The experimental results prove the superiority of the suggested gait analysis method in accurately predicting the severity of Parkinson’s disease. The accuracy rates achieved were 96.34% on the H&Y scale and 97.14% on the UPDRS scale.

Further, to investigate the impact of AI on multimodal data, the research focused on using PET images from PD patients and healthy controls to diagnose PD. Model interpretability techniques, such as LIME and SHAP, help identify regions of interest and factors influencing classification decisions, aiding in understanding the underlying pathophysiology of PD and developing biomarkers for disease progression. This will ensure that we understand how models come to their conclusions. Most importantly, the explainable AI techniques known as LIME highlight the most critical regions contributing to the PD classification, yielding a higher accuracy rate of 94.4% using Alex Net for the proposed PD classification method.

Download

Share

COinS