Assessing Parkinson’s Gait via Smartphone Video and AI
Original Title: Deep learning-enabled accurate assessment of gait impairments in Parkinson's disease using smartphone videos
Journal: NPJ digital medicine
DOI: 10.1038/s41746-025-02150-8
Overview
Parkinson's disease affects millions globally, with gait impairments serving as a primary source of disability. Traditional assessment relies on the Unified Parkinson’s Disease Rating Scale, which is often subjective and lacks the sensitivity to detect minor changes. This study introduces a deep learning framework that utilizes videos recorded by a single smartphone from lateral perspectives to evaluate gait. The system employs a Siamese contrastive network architecture to fuse information from both sides of the body. In testing, the model achieved a micro-average area under the receiver operating characteristic curve of 0.87 and an F1 score of 0.806 for predicting disease severity. These results are comparable to the performance of three clinical specialists, who showed an average error rate of 0.19 versus the model's 0.20. Furthermore, the framework identified medication-induced changes with 73.68% precision. It successfully detected subtle improvements in patients whose clinical scores remained unchanged, demonstrating a higher resolution than standard rating scales.
Novelty
The technical contribution lies in the design of a Siamese contrastive network that integrates dual-perspective lateral videos, allowing for a comprehensive analysis of bilateral symmetry and gait dynamics. Unlike previous methods that often rely on frontal views or multiple fixed cameras, this approach uses a single mobile device. The researchers implemented a modification algorithm to correct leg keypoint misidentifications that occur when limbs cross in lateral footage. Another distinct feature is the interpretability provided by a dual maximum gradient-weighted class activation mapping method. This allowed the extraction of both traditional markers, such as arm swing amplitude (correlation of -0.64), and novel digital biomarkers. Specifically, the average linear velocity of the ankle emerged as a highly sensitive indicator with a correlation of -0.66 to disease severity. The study also introduced a fine-granular assessment score that integrates predicted probabilities and confidence levels to identify medication responses that fall between the coarse increments of clinical scales.
Potential Clinical / Research Applications
This technology facilitates objective, longitudinal monitoring of Parkinson's disease in domestic environments, reducing the need for frequent hospital visits. In clinical research, the framework can serve as a sensitive tool for evaluating the efficacy of new disease-modifying therapies, potentially detecting therapeutic signals that traditional scales might miss. The identification of specific biomarkers, such as neck and head acceleration, allows for personalized medication adjustments based on individual motor responses. Beyond Parkinson's disease, the methodology could be adapted to assess motor impairments in other neurological conditions, including stroke recovery and Alzheimer's disease. The availability of a simplified online assessment system further lowers the barrier for clinical adoption and data collection across diverse populations.