A Hybrid Approach Using 2D CNN and Attention-Based LSTM for Parkinson’s Disease Detection from Video

The development of a deep learning-based approach for Parkinson’s Disease (PD) detection presents a promising solution to enhance diagnostic precision and consistency. The current diagnostic process intensely relies on subjective clinical judgment, resulting in changeable accuracy influenced by clinician skills. To solve this limitation, we present a hybrid approach using 2D CNN and attention-based LSTM network that takes video recordings as input, basically eliminating the need for wearable sensors and expediting the diagnosis process. We are particularly interested in assessing parkinsonian gait, a recognizable distinct indicator of PD, using a pre-trained Convolutional Neural Network (CNN) paired with an attention mechanism (AM). The CNN extracts relevant indicators of gait abnormalities, transmitted afterwards through an attention layer to a Long Short-Term Memory (LSTM) network to improve classification accuracy and detection performance. Empirical results demonstrate the effectiveness of our method, achieving a 92.05% accuracy in distinguishing parkinsonian from non-parkinsonian gait patterns in both training and testing datasets. These findings underscore the potential of our approach as a crucial tool for PD detection and diagnosis.