Parkinson's disease (PD), the second most common neurodegenerative disorder, affects patients and caregivers worldwide. There is a growing need for technologies that use portable wearable sensors to monitor and assess patients using artificial intelligence (AI). Nevertheless, many state-of-the-art AI recognition systems are black-box models, making it difficult to understand their internal logic and leading to mistrust in medical applications. In this paper, we propose an eXplainable hybrid attention convolutional neural network for Parkinson's disease monitoring and evaluation (PDXNet) based on real-world PD data from the University of Rochester. Specifically, we automatically extract resting tremor segments and use a hybrid-attention CNN to evaluate tremor presence (AUC > 0.93). Moreover, saliency maps and SHapley Additive exPlanations (SHAP) are used to interpret the model by visualizing feature contributions for human experts to assess the model's internal logic. Ultimately, the feature contributions obtained align with human intuition. PDXNet demonstrates high accuracy in assessing resting tremors in PD and provides interpretability to users. It offers a trustworthy technical solution for PD monitoring using wearable sensors and AI.
Zhang et al. (Thu,) studied this question.