Mobile Phone-Based Digital Biomarkers Empowered by Knowledge Distillation for Diagnosis of Parkinson's Disease

Tongyue He, Junxin Chen, Chi Lin, Wei Wang

Abstract: Mobile phones have evolved from basic communication tools to feature-rich mobile devices. These ubiquitous and portable devices, equipped with inertial sensors and high-speed network access, create opportunities for remote health monitoring, especially for movement disorders such as Parkinson’s disease (PD). Inertial sensors (gyroscopes and accelerometers) endow smartphones with a natural ability to monitor movement disorders. Based on this, we develop a novel vision-based time-series feature augmentation framework for remote diagnosis and severity grading of PD using mobile phone walking records. Specifically, preprocessed time-series data is encoded into RGB images for the teacher model, while the time-series data is input into the student model, with the teacher guiding the student’s learning. The teacher model is based on MobileNetV2 and incorporates spatial and channel relation-aware attention mechanisms to capture important features and filter out irrelevant information. The inter-modal feature fusion module combines attention and CNN to emphasize both global and local features. The student model utilizes a simple CNN to directly extract features from time-series data and perform classification. For the three-level classification task, the teacher model achieves accuracies of 0.887, 0.886, and 0.896 across the three datasets, while the distillation student model reaches 0.779, 0.828, and 0.827, generally surpassing state-of-the-art algorithms.

External IDs:dblp:journals/tmc/HeCLW25