Movement disorders are a group of neurological conditions caused by dysfunction of the nervous system that affect the speed, quality, and ease of movement. They can be broadly classified into hyperkinetic (excessive movement) and hypokinetic (slow movement) types. The phenomenology of movement disorders is critical since it acts as a clue to the etiological diagnosis and symptomatic treatments. Traditionally, the assessments of the phenotype are carried out using the scales by the clinical raters. In addition to inter-rater inconsistency, which is always an issue, scale evaluations are time-consuming and limited by clinical settings, all of which affect their objective assessment of symptoms and treatment efficacy, particularly in real-world scenarios requiring continuous monitoring of the disease. Moreover, lots of useful information might be missed in the data collecting by digital materials of video and audio recordings, neuroimaging, electrophysiological signals, wearable sensors if they are analyzed by traditional methods.Artificial intelligence (AI) based methods, especially the novel data driven models using machine learning and deep learning, showed potentials for interpreting complex data into quantifiable metrics, which have revolutionized medical data analysis and image processing, assisting in deep phenotyping (1), diagnosis (2, 3), identifying digital biomarkers(4), remote screening (5) and monitoring in movement disorders (6). Chen et al. proposed a method to classify PD from healthy controls, using voice recordings of the subjects. In their study, three pre-trained deep learning models (convolutional neural networks) were used to extract features from the spectrograms converted from the speech segments, followed by feature fusion.The fusion of 2 models achieved the highest accuracy of 95.56% and an AUC of 0.99, better than the best single model (92.73% accuracy).These methods provide a low-cost, non-invasive and a convenient tool for PD diagnosis. They also emphasized the importance of multimodal framework and multi-model integration in the proposed mode, since the different data types and different algorithms attained better model performance.Non-motor symptoms were important but seemed less enrolled as phenotype in the AI analysis. Here, Serbee et al. investigated the association between hypomimia and sialorrhea in PD by AI base video analysis of a 10second videos of happy facial expressions and found the moderate correlations between sialorrhea rating scales and features derived from facial regions. The videos were analyzed using AI algorithms to extract key facial landmarks which were then processed into 20 quantitative features representing mouth, eyes, and combined facial regions. Principal Component Analysis, Canonical Correlation Analysis, and bootstrapping were applied. Their study indicated the video-based assessment tools may also be extended to screen for non-motor symptoms, which broadened the scope of digital phenotyping in PD.AI based approach quantifying motor symptoms first characterized the kinematic effects of levodopa on PD related motor features and found 3 kinematic domains of speed, consistency, timing/scale improved by levodopa, reported by Lange and colleagues in 2025, describing the effects of treatments unable to be detected by conventional scales (7). The unqualified or inconsistent data, patients' privacy protection, poor model generalization, lack of real-world validation or external validation, blackbox effect, might be the challenges and concerns of AI in movement disorders and should be improved in the future (6). Finally, it is important to point out that AI serves as an auxiliary tool for diagnosis and management in movement disorders, but it could not replace the role of a physician currently (8).
Sun et al. (Wed,) studied this question.