AI-Powered Smile Analysis: A Breakthrough in Parkinson’s Disease Screening

Highlights

• AI models analyzing smile videos achieved high accuracy (up to 87.9%) in detecting Parkinson’s disease (PD).
• Largest dataset to date: 1,452 participants, 391 with PD, enabling robust machine learning.
• Screening method showed strong generalizability across North American and Bangladeshi populations.
• Potential for scalable, remote, and low-cost PD screening where clinical access is limited.

Study Background and Disease Burden

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor and non-motor symptoms, including tremor, rigidity, bradykinesia, and the often-overlooked sign of reduced facial expressiveness (hypomimia). Early and accurate diagnosis is critical for optimizing care, but PD is frequently underdiagnosed, especially in regions with limited neurological expertise. Traditional diagnosis relies heavily on in-person clinical assessment, which is not feasible for many populations due to geographic, socioeconomic, or resource constraints. This unmet need for accessible, scalable screening tools has driven research into digital and artificial intelligence (AI)-enabled diagnostic approaches.

Study Design

This study, published in NEJM AI (2025), reports on the largest video-based facial expression dataset for PD screening, comprising 1,452 unique participants from multiple countries, primarily North America and Bangladesh. Of these, 391 had PD (300 clinically diagnosed; 91 self-reported). Recruitment was achieved through social media, email outreach, PD research registries, in-person clinics, PD wellness centers, and high-risk identification in Bangladesh.

Participants used an online tool to record themselves mimicking three facial expressions—smile, disgust, and surprise—either at home or in clinical settings. For this analysis, the focus was on smile videos. Advanced computer vision algorithms extracted facial landmarks and action unit–based features to quantify hypomimia. Machine learning models were trained to distinguish PD from non-PD based on these features. Model generalizability was tested using external datasets from a U.S. clinic and a Bangladeshi cohort.

Key Findings

The AI models demonstrated the following performance:

• Internal 10-fold cross-validation (smile videos): Accuracy 87.9 ± 0.1%, AUROC 89.3 ± 0.3%, Sensitivity 76.8 ± 0.4%, Specificity 91.4 ± 0.3%, PPV 73.3 ± 0.5%, NPV 92.7 ± 0.1%.
• U.S. clinic test set: Accuracy 80.3 ± 1.6%, AUROC 83.3 ± 1.4%, Sensitivity 80.0 ± 2.5%, Specificity 80.5 ± 2.0%.
• Bangladesh test set: Accuracy 85.3 ± 1.4%, AUROC 81.5 ± 1.8%, Sensitivity and specificity remained high, but PPV decreased to 35.7 ± 4.8% due to lower disease prevalence.

No significant performance biases were observed across sex and ethnic subgroups, except in the Bangladeshi test set, where diagnostic accuracy was higher in female participants.

These results indicate that smile-based hypomimia detection via AI is robust across diverse settings and populations. The high negative predictive value (NPV) suggests utility as a broad screening tool to rule out PD in low-prevalence settings.

Expert Commentary

AI-driven digital phenotyping—particularly the quantification of hypomimia—represents a promising frontier in neurology. The use of a simple smile, captured via a smartphone or webcam, as a diagnostic biomarker is compelling due to its universal accessibility and minimal training requirements for users. The study’s robust multi-site dataset and external validation strengthen its clinical relevance.

However, several limitations warrant attention:

• Self-reported PD cases (23% of the PD group) could introduce diagnostic misclassification.
• PPV is modest in low-prevalence settings, indicating that positive screens require confirmatory clinical evaluation.
• Performance differences by sex in the Bangladesh cohort suggest a need for further research into cultural or phenotypic expression variations.
• The model’s current focus is on screening, not diagnosis, and does not address PD subtypes or disease staging.

From a public health perspective, such AI tools could bridge gaps in neurological care, facilitate earlier specialist referrals, and support large-scale epidemiological surveillance, especially in underserved areas. Integration into telemedicine platforms or community screening initiatives could transform pathways to PD diagnosis.

Conclusion

AI-enabled analysis of smile videos offers a feasible, accurate, and scalable method for screening Parkinson’s disease across diverse populations. While not a substitute for expert clinical diagnosis, this technology can expand access to screening, triage high-risk individuals, and reduce diagnostic delays, particularly in resource-limited settings. Future directions include refining algorithms for greater specificity, integrating multimodal digital biomarkers, and evaluating real-world implementation outcomes.

References

• AI-Enabled Parkinson’s Disease Screening Using Smile Videos. NEJM AI 2025;2(7) DOI: 10.1056/AIoa2400950
• Dorsey ER, et al. The Emerging Evidence of the Parkinson Pandemic. J Parkinsons Dis. 2018;8(s1):S3-S8.
• Postuma RB, et al. MDS clinical diagnostic criteria for Parkinson’s disease. Mov Disord. 2015;30(12):1591-1601.