What is the smile as a digital biomarker for Parkinson's?

AI analyzes brief smile videos to extract facial movement features (mouth opening, muscle activation, eyebrow motion) that act as a digital biomarker indicating Parkinson's-related hypomimia and motor abnormalities.

Why is accessible Parkinson's screening needed?

Accessible screening is needed because Parkinson's is the fastest-growing neurodegenerative disorder, motor symptoms appear late, and many regions lack specialists; remote, affordable tools enable earlier detection and wider population reach.

How was the smile-based AI study conducted?

The study enrolled 1,452 participants who recorded smile, surprise, and disgust videos; facial features were extracted with OpenFace and MediaPipe, and HistGradientBoost models were trained to distinguish Parkinson's using smile metrics.

What diagnostic accuracy did the smile model achieve?

Internal validation showed 88% accuracy, 77% sensitivity, 91% specificity, and 93% negative predictive value. External tests in the U.S. and Bangladesh reported 80–85% accuracy, comparable to some specialists.

What are the clinical implications of smile-based screening?

Potential uses include remote preliminary screening, prioritizing neurology referrals, and complementing telemedicine or monitoring programs; it can aid early intervention but does not replace formal clinical diagnosis.

Are there known biases or limitations in the smile screening approach?

Limitations include dataset composition (clinical and self-diagnosed cases), variable recording conditions, and the need for larger, diverse validation; potential biases and generalizability require further assessment before deployment.

Early screening for Parkinson’s disease with AI: the smile as a digital biomarker

PhD candidate Marta Arbizu Gómez explains in this article the recent study “AI-Enabled Parkinson’s Disease Screening Using Smile Videos”, which presents the smile as a digital biomarker for Parkinson’s.

Why do we need accessible screening methods for Parkinson’s?

The Parkinson’s disease (PD) is the fastest-growing neurodegenerative disorder worldwide, affecting more people each year and creating a profound impact on patients, their families, and healthcare systems.

Current diagnosis depends on motor symptoms such as tremor and rigidity, which usually manifest in advanced stages, when there is already significant loss of dopaminergic neurons. Added to this is the difficulty of access to specialized neurologists, especially in rural areas or in low-resource countries.

Having a simple, affordable tool that can be applied remotely would allow:

Detect early signs of the disease and enable early interventions.
Expand the reach of screening in populations with limited access to specialists.
Reduce costs and logistical barriers, leveraging everyday devices like smartphones or webcams.

How was the research conducted?

A recent study published in NEJM AI proposes using artificial intelligence (AI) to analyze smile videos to identify potential Parkinson’s cases.

The researchers recruited 1,452 participants from different parts of the world, including the U.S. and Bangladesh. Of them, 391 people had Parkinson’s disease (300 clinically diagnosed and 91 self-diagnosed).

Each person recorded short videos performing three facial expressions: smile, surprise, and disgust. From these videos, key facial features were extracted using advanced tools such as OpenFace and MediaPipe. Features of interest included mouth opening, facial muscle activation, and eyebrow movements.

With the obtained information, machine learning models (HistGradientBoost models) were trained to distinguish between people with and without Parkinson’s, focusing on smile analysis.

Subscribe
to our
Newsletter

What do the key results reveal?

The results were especially promising for the following reasons:

Overall accuracy in internal validation: 88%.
Sensitivity (ability to correctly detect PD cases): 77%.
Specificity (ability to correctly exclude non-PD cases): 91%.
Negative predictive value (NPV): 93% (very useful for ruling out the disease).

In external tests conducted in the U.S. and Bangladesh, accuracy ranged between 80% and 85%, remaining competitive with the accuracy achieved by some specialists.

Figure 1 Comparative bar chart of the main diagnostic metrics, showing the model's performance in internal validation, the U.S., and Bangladesh. — Figure 1. Comparative bar chart of the main diagnostic metrics, showing the model’s performance in internal validation, the U.S., and Bangladesh.

An interesting finding is that the smile turned out to be the most discriminative expression, outperforming surprise and disgust. Additionally, the analysis did not show significant sex or ethnicity biases in most datasets.

Therefore, the use of facial expressions in Parkinson’s screening may represent a useful tool.

Figure 2 Comparative diagram showing the model's performance (AUROC) for each facial expression: smile, disgust, and surprise. — Figure 2. Comparative diagram showing the model’s performance (AUROC) for each facial expression: smile, disgust, and surprise.

What are the implications for clinical practice?

The study demonstrates AI’s potential to transform early Parkinson’s screening. Notable applications include:

Remote and accessible screening: people in rural areas or without access to neurologists could record a simple video from home and receive a preliminary assessment.
Reduce congestion in specialized clinics: prioritize referrals to neurology for cases with a higher likelihood of disease.
Complementary monitoring: although it does not replace a clinical diagnosis, it could serve as additional follow-up in telemedicine programs.

How does this advancement relate to NeuronUP?

At NeuronUP we develop stimulation and cognitive rehabilitation programs based on scientific evidence. Incorporating technologies that analyze digital biomarkers such as facial expressivity opens the door to:

Personalization of interventions according to the degree of motor and cognitive impairment.
Early detection, allowing rehabilitation programs to be started earlier and maximize their effectiveness.
Multidisciplinary integration, combining digital assessment and cognitive stimulation for a more comprehensive and continuous approach.

Conclusion

The smile, a universal and seemingly simple gesture, can become a powerful tool for early detection of Parkinson’s disease thanks to artificial intelligence. This pioneering approach not only brings diagnosis closer to more people but also lays the groundwork for a future where technology and rehabilitation work hand in hand to improve the quality of life of those living with this disease.