Explainable Gait Analysis for Early Detection of Neurodegenerative Diseases Using Unsupervised Clustering Techniques

Neurodegenerative diseases are the most frequent age-related diseases. It is crucial to be able to detect this type of disease at an early stage so that early intervention can take place, improving the quality of life of patients and the loved ones who care for them. One of the tasks recognized by neurologists to detect the onset of neurodegenerative diseases and their relative severity is gait analysis. This work presents a dataset of time-series coordinates of certain body parts such as the head, shoulders, etc. extracted through the use of a camera. The work aims to be able to identify through certain features extracted from these coordinates the presence or absence of neurodegenerative disease. This would allow a less invasive, low-cost, and faster diagnosis. To achieve this goal, explainable artificial intelligence (XAI) techniques have been employed to enhance the interpretability and transparency of the unsupervised clustering models used for gait analysis. Specifically, we leverage unsupervised clustering algorithms such as K-means, k-medoids, Gaussian Mixture Models (GMM), and Agglomerative Clustering to identify patterns in the gait data. Methods: The videos were obtained with a video camera, and through a library of estimated poses, the time series for the joint points were extracted. The data-mining part was performed with an unsupervised paradigm using four clustering algorithms: K-means, k-medoids, GMM, and Agglomerative Clustering. However, unlike traditional black-box models, our approach incorporates XAI methods, such as dimensionality reduction visualization and clusters heatmaps, to provide explanations for the clustering results, enabling better understanding and trust in the diagnostic process. Results: Agglomerative clustering achieved an accuracy of 98% and a sensitivity of 100% in identifying neurodegenerative diseases based on gait patterns. The incorporation of XAI techniques provided valuable insights into the most discriminative features contributing to the clustering results, enabling better interpretation and potential clinical application of the proposed approach.