Machine Learning Aided Kinematic Profiling of Reaching Movements Separates Spinocerebellar Ataxia type 12 and Essential Tremor.

Upper limb postural tremor is common in both Spinocerebellar Ataxia type 12 (SCA12) and Essential Tremor (ET). In the early stages, their clinical presentation can be similar, leading to potential misdiagnosis. This study aimed to develop and evaluate a low-cost device and methods to quantify 2D reaching kinematics objectively, with the goal of improving differentiation between ET and SCA12 using machine learning (ML) models. We recruited 48 participants, including 16 SCA12 patients, 16 ET patients, and 16 healthy controls (HC). Clinical severity was scored using TETRAS and SARA scales. A Raspberry Pi-based device was used to record upper limb reaching during a button-press task. The videos were processed with DeepLabCut and six 2D kinematic parameters were extracted via MATLAB scripts. We expanded the dataset using Generative Adversarial Networks (GAN). Classification was performed using multiple ML algorithms, with and without GAN-based data augmentation. Both patient cohorts showed significant alterations in kinematic features compared to HC, with SCA12 demonstrating more pronounced abnormality. Kinematic parameters including maximum reaching speed, endpoint precision, and time fraction to peak speed emerged as potential markers for differentiating ET and SCA12. The ML models demonstrated high classification performance in distinguishing HC from patients (CA 88.9%) and SCA12 from ET (CA 83.3%) using original data. The device-guided reaching kinematics provide objective markers of motor impairment. Combined with ML, this approach could enhance diagnostic accuracy; it also has potential for remote monitoring and as an outcome measure in clinical trials.