Introduction

Parkinson’s disease (PD) is a neurodegenerative disorder defined by the loss of dopamine-secreting neurons and resulting motor symptoms, including bradykinesia, gait instability, and resting tremor, as well as a range of non-motor features1. With no approved disease-modifying therapies currently available, treatment is symptom-directed. However, diverse symptom profiles across patients, day-to-day fluctuations, and variable responsiveness to medication make this approach challenging2,3. For example, the medication levodopa may reduce a patient’s hand tremor but worsen their fall risk4. Detailed, frequent symptom assessment is therefore crucial to support tailored treatment plans.

Symptom assessment today often relies on the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS)5,6. Although comprehensive, the scale has limited sensitivity in early disease, requires trained administrators, and raises concerns regarding inter- and intra-rater reliability5,7,8,9,10. Further, its ordinal scoring is insufficiently granular to support individualized therapy. Gait impairment is captured by one dedicated item employing a five-point scale and supplemented with similar items for posture and freezing. Clinically meaningful detail, such as the presence of appendicular versus axial rigidity, cannot be extracted, though the former is less associated with functional impairment and more responsive to levadopa11,12.

Objective gait assessment is a promising complementary biomarker. Gait impairment strongly affects quality of life and remains poorly responsive to many therapies2,5,10,11,13. More precise characterization of its fluctuating domains could help identify effective interventions, such as physical therapy targeting specific deficits, and improve the prediction of a medication’s functional impact7,10,14. In addition, subtle gait changes can precede clinical PD diagnosis by years15,16. Identification of these high-risk individuals is crucial for clinical trials, as they may be more responsive to future disease-modifying therapies17,18,19.

Novel biomarkers from smartphone videos

In “Deep learning-enabled accurate assessment of gait impairments in Parkinson’s disease using smartphone videos,” Han et al. develop a system for objective gait assessment20. Using side-view, smartphone videos of patients walking, they apply a deep learning framework to extract joint movements that predict gait impairment severity. The model’s predictions closely track the consensus MDS-UPDRS gait scores of three specialists (F1 score 0.806), with disagreements limited to one-point score differences and concentrated in cases where the individual expert ratings were also inconsistent, potentially reflecting clinical uncertainty or MDS-UPDRS limitations (model error rate 0.20 versus expert error rate 0.19). The authors then test whether the model can detect medication-related changes in gait, distinguishing on- versus off-medication states using consensus MDS-UPDRS scores and a more granular gait subscore as reference standards. In this setting, the model alone achieves approximately 74% accuracy, outperforming each individual clinician scoring the same videos without assistive analytics.

Beyond classifying disease severity, the model produces interpretable outputs highlighting which movement features drive its predictions, a property that could build clinician trust in its recommendations. The system also surfaces quantitative features—such as ankle and head velocity—that are more strongly associated with disease severity and on/off-medication status than some traditionally emphasized measures, including arm swing10. These findings show that decomposing gait into subdomains yields more predictive, multi-feature assessments than a single summary score. Such granularity may promote more effective patient care by improving sensitivity to medication effects and focusing rehabilitative and other interventions toward gait features most strongly associated with functional impairment.

Addressing access to care

Smartphone-based gait assessment offers not only greater predictive capabilities but also a pragmatic means to improve access to care through remote monitoring and decentralized care models. In the United States, only half of patients with PD regularly see a neurologist, and just 9% see a movement-disorder specialist, despite evidence that specialized care improves outcomes21,22,23. Meanwhile, the number of individuals living with PD is expected to roughly double by 2040, exacerbating the gap between demand for neurologists and the available workforce22,24.

Prior work has shown that video-based analysis can support objective gait assessment, but Han et al.20 uniquely demonstrate that simple smartphone-recorded videos are sufficient to power this approach5,25. This technology could thereby extend access to specialty care if incorporated into telehealth visits or asynchronous monitoring. However, its greatest impact may come through shifting routine gait assessment from specialty settings to primary care practitioners and general neurologists, who are more accessible contacts for patients with PD22.

Implementation challenges

While Han et al. focus on a video-based approach, wearable devices for PD symptom monitoring are also well validated26,27,28. A central question for the field is how complementary modalities should be deployed together in clinical care.

Wearables, such as pressure-sensing insoles or wrist inertial measurement units, offer continuous monitoring representative of typical function and day-to-day variation, while maintaining or exceeding the sensitivity of the MDS-UPDRS3,10,11,26,29,30,31. Yet few clinicians report incorporating wearable outputs into decision-making, possibly reflecting workflow barriers, uncertainty around unfamiliar metrics, and limited evidence regarding impact on clinical outcomes or quality of life3,9,28,32,33,34. Patients also report discomfort, stigma, and frustration with wearable devices, and adherence can be poor11,14.

Video-based approaches address several of these barriers by fitting more naturally into existing workflows5,7,26. Smartphone videos can be captured during routine in-person or telemedicine visits without requiring patients to purchase or wear bothersome hardware, and clinicians may be more comfortable adopting deep learning tools that augment filmed MDS-UPDRS examinations than interpreting home-monitoring data7,22. However, video analysis is episodic, limiting insight into at-home function and constraining usefulness for smart interventions such as deep brain stimulation or infusion pumps. Video quality is another constraint — Han et al. excluded patients wearing baggy clothing and those requiring assistance with walking, although a third of patients with PD use mobility devices, because these features interfered with pose estimation25,35,36. Moreover, all videos were obtained by the research team. The feasibility of patient-led filming remains unestablished.

Navigating these points will require a staged research agenda that begins with external validation, or proof of performance across diverse patient groups. Han et al. have not yet conducted such studies, and, for now, expansion of the authors’ technology into clinical care is theoretical. Subsequent trials should define clearly specified use cases, including by whom tools will be used, at what intervals, and in support of which clinical decisions, with pragmatic metrics such as the proportion of patient-recorded videos or wearable data streams that are analyzable and the influence of new data on medication adjustments.

Conclusion

Han et al. demonstrate that deep learning applied to simple smartphone videos can approximate specialist gait ratings, distinguish medication effects, and reveal interpretable movement features. Alongside wearable technologies, the authors’ approach paves the way to personalized treatment and expanded access to care for patients with PD. Validation in broader populations and evaluation of clinical impact will determine how to integrate this approach into everyday care.