Deep bispectral image analysis for imu-based parkinsonian tremor detection

Tremor is the most common motor symptom of Parkinson's Disease (PD) that deteriorates life quality of patients, with early detection being crucial for inhibiting progression of the disease. Widely available commercial devices have enabled continuous data collection and, thus, methods that contribute towards detection of PD symptoms in-the-wild, are of great importance. In this study, we opt for a method to automatically detect PD tremor using Inertial Measurement Unit (IMU) data captured passively via smartphone, during the user's daily phone calls. To that end, the DeepBispecI model is proposed, where the IMU data are subjected to a Bispectral analysis, resulting in third-order spectrum images that are subsequently fed to a Convolutional Neural Network (CNN). DeepBispecI was applied on IMU data from 31 PD patients and 14 healthy controls, resulting in accuracy, sensitivity, specificity and F1 of more than 95%. This indicates that highly accurate detection of PD tremor episodes is feasible by using data that are collected in-the-wild, thus promoting development of applications that can be utilized for continuous monitoring of PD tremor symptoms through the user-smartphone interface.