Quantitative Assessment of Facial Expression Asymmetry in Parkinson’s Disease

doi:10.21203/rs.3.rs-5315631/v1

Hypomimia, characterized by reduced facial expression, is a cardinal feature of Parkinson's Disease (PD). However, unlike limb asymmetry in PD, facial asymmetry has been less explored. Here, we explore possible subtle hemihypomimia in PD using Artificial Intelligence (AI) and image processing techniques. After video preprocessing facial expression videos from 102 PD subjects and 97 healthy controls (HCs), asymmetry index values across facial landmarks were calculated for each frame. Dynamic features were extracted and used in machine learning models to differentiate between PD and HCs, achieving 91.4% accuracy. PD subjects showed greater facial asymmetry, particularly around the eyebrows (P = 0.01) and mouth (P = 0.04), and those with asymmetric limb Parkinsonism exhibited less facial mobility on the more affected side (P = 0.001). These findings support the presence of facial expression asymmetry in PD, particularly during expressions of happiness, and suggest its potential as a clinical digital biomarker.

Health sciences/Diseases/Neurological disorders/Neurodegenerative diseases/Parkinsons disease

Biological sciences/Neuroscience/Diseases of the nervous system/Parkinsons disease

Health sciences/Biomarkers/Diagnostic markers

Parkinson disease (PD) is typically associated with asymmetric motor onset in the limbs, a characteristic that persists throughout the disease course^1,2. Approximately 85–95% of subjects with PD have asymmetry in limb Parkinsonism^3–6. Axial symptoms, which affect the body's midline, include difficulties in swallowing, balance, and gait, and usually emerge later in the disease course. Unlike appendicular symptoms, which often respond well to dopaminergic therapy, axial symptoms show limited response⁷.

Hypomimia, a reduction in facial expression, is a common and early motor manifestation in PD⁸. Hypomimia may be a relatively unique characteristic, as it can show a favorable response to levodopa, despite being considered a symmetric axial sign because of its midline location⁹. However, is hypomimia really symmetrical? Facial musculature, especially in the lower face, is innervated separately by the contralateral motor cortex on each side. There are several anecdotal reports of subjects with PD who manifest 'hemihypomimia', where reduction of facial expression occurs only unilaterally^10–12. Nevertheless, facial asymmetry has not been adopted as a diagnostic criterion for PD, possibly due to the challenges in detecting subtle facial muscle movements. This may also explain why asymmetry in hypomimia has not been thoroughly studied ¹³.

Recent advances in artificial intelligence (AI) and computer vision technology have made facial expression analysis more accessible, enhancing the accuracy of detecting subtle asymmetries in facial muscle movement that might be missed by the naked eye. For example, AI approaches have been used in treatment planning and preoperative evaluation for orthodontic or maxillofacial surgeries to preserve aesthetic facial symmetry^14,15. They have also been employed to quantify facial asymmetry from 2D videos for evaluating facial paralysis¹⁶.

In this study, we investigated the asymmetry of hypomimia in subjects with PD using quantitative computer vision techniques. We developed indices to quantify facial asymmetry from videos capturing various emotional facial expressions of the participants and compared these indices between PD subjects and healthy controls. We applied several machine learning algorithms to distinguish PD subjects from healthy subjects based on both static and dynamic features derived from the indices, without referencing limb symptoms. We tested the hypothesis that facial asymmetry is more significant in PD compared to controls. This approach aims to enhance the understanding of facial asymmetry in PD and to explore its potential as a diagnostic tool.

Baseline characteristics

The study included 199 participants, 102 with PD and 97 HCs. The age of the HC group was younger (52 ± 15 years) compared to the PD group (60 ± 10 years) (P = 0.01) in the UT dataset, which includes 103 participants (25 subjects with PD and 78 healthy controls). The UBC dataset comprised 96 participants (77 PD subjects and 19 HCs), and intentionally did not collect demographic information, including age and disease duration, because of the center's study aims to minimize the burden on subjects for data entry and to focus on assessing Parkinsonism using solely the video data (Table 1)

Table 1

Baseline demographics
	PD (n = 102)	healthy controls (n = 97)	P value
Dataset			-
UT dataset	25	78
UBC dataset	77	19
Age^*	60 ± 10	52 ± 15	0.01
Sex			0.25
Male	71 (70)	60 (62)
Female	31 (30)	37 (38)
Disease duration^*, years	6.12 ± 6.47	-	-
Hypomimia^**			-
0	7 (9)	-
1	12 (16)	-
2	35 (46)	-
3	15 (19)	-
4	8 (10)	-
Finger tapping^** symmetry			-
Symmetric (left = right)	40 (51)	-
Asymmetric (left ≠ right)	37 (49)	-
Finger tapping^** laterality			-
Right dominant	15 (40)	-
Left dominant	22 (60)	-
Data are shown in either mean ± standard deviation or n (%).
^*Data from UT dataset only.
^**As assessed by MDS-UPDRS items (3.2. facial expression and 3.4. finger tapping) in UBC dataset only.

Comparing the asymmetry indices between PD and HCs

We compared the range of the Euclidean facial asymmetry indices of happy facial expression videos (maximum - minimum value from a single video) between PD and HCs across the 6 facial ROIs using Mann-Whitney U test. The ranges of indices were significantly larger in PD compared with HCs, particularly in the eyebrows (PD, 1.47 ± 1.36; HCs, 1.04 ± 0.58; P = 0.01) and mouth (PD, 1.23 ± 1.21; HCs, 0.79 ± 0.42; P = 0.04) area (Table 2). In addition, though not statistically significant, the ranges of asymmetry indices for the eyes, nose, and facial outline were marginally higher in the PD group compared with HCs (P = 0.06, P = 0.05, P = 0.05, respectively) (Table 2).

Table 2

Comparison of the range of facial asymmetry indices of happy facial expression videos between PD and healthy controls
Facial ROIs	Healthy Controls (n = 97)	PD (n = 102)	P value
Whole face	1.84 ± 1.13	2.84 ± 3.17	0.10
Eyebrows	1.04 ± 0.58	1.47 ± 1.36	0.01
Eyes	1.05 ± 0.64	1.54 ± 1.55	0.06
Nose	0.84 ± 0.48	1.28 ± 1.35	0.05
Mouth	0.79 ± 0.42	1.23 ± 1.21	0.04
Facial outline	0.95 ± 0.59	1.54 ± 1.70	0.05

Differentiating PD and HC using Facial Asymmetry Indices

Next, we explored whether PD and HC can be distinguished accurately based solely on the asymmetry indices of facial expressions. Specifically, we aimed to investigate the contribution of the static or dynamic nature of the input features. The static features were the Euclidean and vertical facial asymmetry indices of the 6 facial ROIs of each frame. The dynamic features included the basic time-series features such as mean, variance, standard deviation, amplitude, root mean square (RMS), and the CATCH22 time-series features. We prepared six distinct sets of input features: i) static features from a single frame randomly selected from the neutral facial expression video, ii) static features from a single frame identified as the apex intensity of the happy facial expression video, iii) static features from 5 randomly extracted frames from the happy facial expression video, iv) static features from 20 randomly extracted frames from the happy facial expression video, v) basic time-series features of the asymmetry indices from the happy facial expression video, and vi) both basic and CATCH22 time-series features of the asymmetry indices from the happy facial expression videos.

Results, as shown in Table 3, revealed a significant disparity in classification accuracy depending on the input and the features used. The Logistic Regression models that demonstrated the highest accuracy among all the classifiers tested are reported. Among the six different settings of input features, the model which included both basic and CATCH22 time-series features of the asymmetry indices achieved the greatest average accuracy of 91.4% in cross-validation (Table 3). The same model achieved a precision rate of 90.0%, a recall of 90.0%, and an F1 score of 90.0% on the validation dataset.

Table 3

Results of the classification using different input features
Source	Input features	Accuracy(STD) (%)	F1 score (%)	Precision (%)	Recall (%)
Neutral expression, single frame	Static	59.0(\(\:\pm\:\)0.09)	41.0	40.2	43.2
Happiness, single apex frame	Static	69.4(\(\:\pm\:\)0.08)	75.2	78.8	75.0
Happiness, 5 frames	Static	65.6(\(\:\pm\:\)0.06)	67.2	67.6	67.0
Happiness, 20 frames	Static	73.5(\(\:\pm\:\)0.05)	68.7	68.7	68.8
Happiness, whole video	Dynamic, basic	77.8(\(\:\pm\:\)0.08)	82.5	82.7	82.5
Happiness, whole video	Dynamic, both basic and CATCH22	91.4(\(\:\pm\:\)0.04)	90.0	90.0	90.0

The model using the asymmetry indices from a single image of neutral facial expression resulted in the lowest mean accuracy of 59.0%. When using the indices from a single image but with a happy facial expression at its peak intensity, the accuracy rose to 69.4%. This discrepancy of accuracy between the resting state and active state of facial muscular activity shows greater asymmetry of facial expression, in other words, asymmetric hypomimia, in subjects with PD. Classification based on sets of 5 frames and 20 frames extracted from happy facial expression videos, as well as basic dynamic features from the entire video, yielded increasing accuracies of 65.6%, 73.5%, and 77.8%, respectively, which further improved to 91.4% when the CATCH22 time series features were incorporated. These results suggest that the facial asymmetric hypomimia is more detectable in the dynamic features compared to the static features.

Aligning the dominant side of the limb and face Parkinsonism

To verify the validity of our hypothesis that hypomimia in PD subjects is asymmetric, we investigated whether the dominant side of hypomimia aligns with the dominant side of limb Parkinsonism. This experiment included 37 participants who exhibited asymmetry of limb Parkinsonism as observed through their finger-tapping tasks (MDS-UPDRS part III finger tapping item difference between right and left side ≥ 1) To measure facial muscle activity, we calculated the average of the summation velocity of changes in distances between facial landmarks and the tip of the nose during happy expression videos. The side with a lower value is considered having slower facial muscle movement than the other side. We compared these facial mobility measures between the dominant and non-dominant symptom side of the limbs. Student's t-test revealed that facial mobility measures are smaller in the side where the limb symptom is dominant (P = 0.002) (Fig. 1). The method successfully determined the affected side of the body with 73% accuracy.

Facial Asymmetry Across Various Emotions and Facial ROIs

Next, we explored whether the asymmetry of facial expression varies among different emotional expressions across different facial regions. The asymmetry indices were extracted and normalized in the 6 facial ROIs (entire face, facial outline, eyebrows, eyes, nose, and mouth) of 4 major emotional expressions (happiness, anger, sadness, and disgust).

This experiment revealed notable variations in the asymmetry of facial expressivity linked to PD (Figure 2). In general, the facial expressions of subjects with PD were more asymmetric than HCs in all tested emotions (Figure 2a, 2b). In subjects with PD, a happy facial expression was found to induce the highest levels of facial asymmetry (Figure 2a, 2c) whereas a disgusted facial expression was the most asymmetric emotional expression in HCs. Eyebrows were the facial ROIs with the most pronounced asymmetry in both subjects with PD and HCs, but the asymmetry of eyebrows stood out in happy and disgusted expressions in PD, whereas it stood out the most in sad expressions in HCs.

In this study, we found that facial expression asymmetry was greater in subjects with Parkinson’s disease (PD) compared to healthy controls (HCs), consistent with the lateral asymmetry seen in limb motor symptoms. Importantly, the facial asymmetry indices were distinct enough to differentiate PD subjects from HCs, particularly when analyzing the dynamic aspects of facial expressions. Furthermore, this asymmetry was most pronounced during happy facial expressions compared to other types of expressions.

The predominantly unilateral onset of limb Parkinsonism in idiopathic Parkinson's disease (PD) has been recognized since the earliest descriptions of the condition¹⁷. Such asymmetry is a conspicuous clinical feature of PD that differentiates it from atypical parkinsonism, such as Multiple System Atrophy (MSA)⁶. Although the cause of asymmetrical PD symptom onset is unclear, the basal ganglia circuit's asymmetric involvement with the primary motor cortex is well-established¹⁸. Evidence from imaging studies of presynaptic nigrostriatal integrity such as fluorine-18-labeled fluorodopa (F-dopa)¹⁷ PET scans and autopsy findings show greater neuronal loss and a higher α-synuclein burden on the side opposite to the more affected limbs^19-21.

As the disease progresses, axial symptoms like postural instability and freezing of gait emerge^2,17, and these symptoms are more resistant to levodopa treatment^7,9. While limb bradykinesia is mainly attributed to the depletion of nigrostriatal dopaminergic neurons, axial symptoms may be linked to non-dopaminergic mechanisms, such as dysfunction in cholinergic pathways ^7,9,22.

Hypomimia is a distinctive feature of Parkinson’s disease (PD) and has been classically considered a symmetric, axial symptom, though it shows partial responsiveness to levodopa. However, facial expression is not merely a motor phenomenon, and likely involves a complex interplay of affective and cognitive mechanisms²³. Our findings on the asymmetry of hypomimia suggest that nigrostriatal dopaminergic degeneration plays a significant role in its pathophysiology. The Parkinson’s Progressive Markers Initiative (PPMI) de novo longitudinal cohort study found that dopaminergic transporter density in the putamen and caudate was significantly lower in PD subjects with hypomimia compared to those without, indicating that hypomimia is linked to dopaminergic loss in the striatum²⁴. Notably, the PPMI study did not differentiate between the left and right sides, and our study did not include dopaminergic imaging quantification or evaluate levodopa responsiveness. Further research exploring the relationship between hypomimia laterality, its levodopa responsiveness, and contralateral dopaminergic transporter density could provide deeper insights into this hypothesis.

Our findings show that facial asymmetry in PD subjects varies depending on the expressed emotion, with the most pronounced asymmetry occurring around the eyebrows and mouth. Studies on facial asymmetry in the general population and among actors have indicated that expressions of disgust and happiness display greater asymmetry, supporting our observations in both PD and healthy control (HC) groups²⁵. Previous studies, based on visual assessments by movement disorder specialists, suggested that facial asymmetry in PD is less prominent in the upper face compared to the lower face, attributing this to the bilateral innervation of the upper face from both cerebral hemispheres^11,13. However, our results do not support this explanation, as we observed more significant asymmetry in the eyebrow region. This result should be interpreted with caution, as eyebrow landmarks are more susceptible to errors, particularly in older adults, where the eyebrows may be thin and blend with the skin tone..

We found that screening for PD using facial asymmetry was more effective when using continuous expression data over a 10-second video (equivalent to 300 frames at 30 frames per second) compared to single images or sets of 5 or 20 frames. This effectiveness improved further when incorporating CATCH22 time-series features, which provide a compact yet comprehensive description of continuously changing behaviors. Clinically, it is recommended to observe facial expressions in a continuous manner, paying attention to eye blinks, lip movements, and natural variations in expression rather than relying on isolated snapshots. The MDS-UPDRS guidelines also instruct evaluators to observe these aspects over a 10-second period while subjects speak or at rest to assess hypomimia²⁶. Since facial expressions change constantly, when assessing facial characteristics, including asymmetry, it is important to focus on dynamic changes.

Our computer vision-based method enabled the detection of hypomimia asymmetry that is not obvious to the naked eye, demonstrating that AI technology can enhance clinical diagnoses involving facial expressions. Various clinical conditions affect facial muscle movement, with hypomimia in PD being but one example, and computer vision techniques have been explored in several of these conditions^27-30. For symmetry assessment, these technologies have been applied in Bell's palsy or maxillofacial surgery planning or pre- and post-surgery comparisons^15,31,32. In PD, research has suggested that AI can differentiate between PD subjects and healthy subjects based on facial expressions^28-30. Currently, the clinical gold standard for assessing the severity of hypomimia is the MDS-UPDRS Part III facial expression item which has been criticized for being not fully quantitative, lacking in the assessment of emotional expressivity, and not accounting for facial asymmetry³³. Expanding our AI tool to dynamically assess facial expression and its symmetry could improve the evaluation of hypomimia severity and treatment responsiveness, as well as broaden its application to other facial movement disorders, such as hemifacial spasm or blepharospasm.

While a strength of our study is its pioneering exploration of facial expression asymmetry in PD subjects using AI techniques, several limitations should be addressed. First, the data collection methods were heterogeneous, though this diversity may have contributed to more robust results. We used two separate datasets with different data collection protocols: one center instructed subjects to mimic expressions shown to them, while the other asked for spontaneous expressions. Facial mimicry and spontaneous expressions engage different, though overlapping, neural pathways, which may influence the degree of facial asymmetry observed. Second, the demographic information collected varied between centers, leading to incomplete data. One center lacked age information despite matching HCs by age, while the other center had age data but included HCs who were significantly younger than the PD subjects. Aging can cause wrinkles and changes in eyebrow color, which may affect the accuracy of facial landmark detection and the performance of our algorithm. Lastly, our study only included emotional facial expressions. Hypomimia involves dysfunction across multiple neural pathways, including motor, affective, and cognitive systems, and affects both emotional and non-emotional facial movements. Future studies using AI-assisted facial expression analysis should control for these factors to provide a more comprehensive understanding of facial asymmetry and related aspects.

In conclusion, our study demonstrates the presence of facial expression asymmetry in PD and its potential to effectively distinguish PD patients from healthy controls using AI and computer vision technology. By showing the value of dynamic facial asymmetry indices in differentiating between these groups, our research advances the understanding of hypomimia and lays the foundation for developing more accurate, non-invasive diagnostic tools that could be applied to a wider range of facial movement disorders in routine clinical practice.

Participants

In this study, 199 participants were enrolled from two distinct datasets: the University of Tehran (UT) dataset and the University of British Columbia (UBC) dataset. The UT dataset comprised 103 participants (25 subjects with PD and 78 healthy controls (HCs)), recruited from the Department of Neurology at Rasoul Akram Hospital, Tehran, Iran. The UBC dataset included an additional 96 participants, with 77 PD subjects and 19 HCs who mainly were the subjects' spouses, recruited from the Pacific Parkinson Research Center at the University of British Columbia, Vancouver, British Columbia, Canada. Inclusion criteria for subjects with PD were i) PD diagnosis confirmed by movement disorder specialists based on the United Kingdom Parkinson’s Disease Society Brain Bank criteria, and ii) age over 30. Exclusion criteria for both PD and HCs were i) any history of facial trauma or surgery, ii) ongoing orofacial symptoms or sequelae of previously diagnosed neurological or neuromuscular disease other than PD, iii) diagnosis of atypical Parkinsonism, or iv) incapability of using the designed applications for facial expression video acquisition. The Institutional Review Boards of each center approved the studies, and informed consent was obtained from all participants.

Video Acquisition

We utilized video recordings of the participants exhibiting neutral facial expressions and four key emotions, including happiness, sadness, anger, and disgust. In the UT dataset, participants interacted with a custom-designed smartphone application installed on their mobile phones. They were presented with videos showing facial expressions corresponding to the four emotions. They were instructed to replicate these expressions while being video-recorded using the phones' selfie cameras. In the UBC dataset, a web-based application was used to demonstrate expressions of the four emotions. They were asked to mimic each specific emotion, eliciting genuine emotional expressions. Their emotion expressions were recorded using webcams of the participants' personal devices or a computer (iMac 24', Apple Inc., CA, USA) located in a private area in the waiting room of the Movement Disorder clinic. The videos were reviewed by a movement disorder specialist and then scored according to the Movement Disorder Society-sponsored UPDRS (MDS-UPDRS) Part III²⁶. The overall flow of the subsequent analyses is depicted in Figure 3a.

Ethical Approvals

The research project involving the UT Dataset was reviewed and approved by the University of Tehran's Ethics Committee (Approval ID: IR.UT.PSYED.REC.140.048), and the UBC Dataset was reviewed and approved by the Chair of the University of British Columbia Clinical Research Ethics Board (Approval ID: H18-03548, approved on 10/12/2023). Informed consent was obtained from all participants prior to conducting interviews and video recordings. Participants were informed of their right to withdraw from the study at any time. All collected data are securely stored, remain confidential, and no personal identifiers, such as participants' names, are mentioned in any research reports.

Assessment of Facial Expression Asymmetry

Preprocessing

For each collected video, we extracted every frame and cropped the face within the frame. To mitigate variability introduced by differing head orientations, cropping and aligning of the face is a crucial step before feature extraction. Head orientation can be modeled by three rotational angles: roll, yaw, and pitch (Figure 3b). Since yaw rotation can impact the measurement of facial asymmetry, all the videos were aligned vertically with respect to the yaw angle. Pitch rotation does not significantly affect the detection of facial asymmetry as it impacts both sides of the face to the same amount. Similarly, roll rotation does not inherently affect the assessment of asymmetry. Nevertheless, we zeroed the roll rotation angle to simplify the subsequent processes. We used the angle between the line connecting the center of each eye and the horizon to rotate the face. The scale factor was computed based on the desired distance between the centers of the eyes, set at approximately 30% of the image width. A rotation matrix was then calculated and applied to the image using the OpenCV library³⁴. This resulted in a face that is both centered and oriented upright, irrespective of the original pose (Figure 3c).

Facial Landmark Detection

Facial landmarks were identified on the aligned faces using the Shape Preserving Facial Landmarks with Graph Attention Networks (SPIGA) algorithm³⁵ It identifies 98 facial landmark information, offering a higher level of precision compared to other models that extract fewer landmarks³⁶. This increased granularity is especially beneficial for accurately capturing the nuances of older faces, where wrinkles and other age-related features can significantly impact facial landmark detection.

Facial Asymmetry Indices

After the face cropping and alignment for each frame, we set six regions of interest (ROIs) among the face for further analyses; entire face, eyes, eyebrows, mouth, nose, and face outline. For each ROI of each frame, we mirrored the landmarks from the left side of the face to the right side using the line that connects the eyes' midpoint and the nose midpoint (Figure 3c). Then, the distances between these mirrored landmarks and the actual landmarks on the other side were calculated in two ways: the Euclidean distances and vertical distances. These two metrics served as the facial asymmetry indices for further analyses. Note that the laterality of mirroring does not influence the outcome, as the distance remains unchanged regardless of which side (right or left) is reflected. We calculated the Euclidean facial asymmetry index as the sum of the Euclidean distances between the mirrored landmarks and the actual landmarks divided by the square root of the sum of the squares of the width and height of the ROI Eq. (1).

Feature Extraction

We extracted several features from the calculated Euclidean and vertical asymmetry indices of each video for further analysis. These features can be categorized into static features and dynamic features. The static features included the mean and range (maximum - minimum) of asymmetry indices.

The dynamic features extracted included basic statistical time-series measures and the CATCH22 time-series features³⁷ of the indices from the entire video sequence. Basic statistical measures for time-series data included the mean, variance, standard deviation, amplitude, root mean square, skewness, and kurtosis. The CATCH22 features are a set of 22 carefully selected statistical and mathematical properties that summarize the essential characteristics of time-series data, enabling quick and efficient analysis of patterns and anomalies without requiring extensive preprocessing or domain-specific knowledge³⁷. These features include measures of central tendency, variability, entropy, and autocorrelation, which help in understanding the dynamics of time-series signals.

Statistical Analyses

For the comparison of baseline characteristics between PD and HC, skewness and kurtosis values were assessed to determine the normality of the distribution. For normally distributed data, chi-square tests and t-tests were performed. For non-normally distributed data, the Mann-Whitney U test and the Kruskal-Wallis test were used.

To select the most effective features, the Recursive Feature Elimination with Cross-Validation method was used. This approach offers a systematic approach to find the most significant features by iteratively removing the least important ones and evaluating model performance through cross-validation. This procedure was enhanced through a grid search technique, to fine-tune the hyperparameters for a range of classifiers, including Support Vector Machines (SVM), Random Forest, Logistic Regression, and the XGBoost classifier³⁸. To ensure the models’ robustness and their capacity to generalize across different data segments, 5-fold cross-validation was adopted.

Data Availability

The raw video recordings of subjects cannot be shared due to compliance with privacy regulations. However, the extracted features from these videos are available from the corresponding author on reasonable request.

Code Availability

The code and scripts used to train model and generate the results reported in this study are available on GitHub at https://github.com/irania/FacialAsymmetry. Any specific dependencies and installation instructions are documented in the repository's README file.

Acknowledgment

This study was supported by the John Nichol Chair in Parkinson’s Research (MJM), Cognitive Sciences & Technologies Council in Iran, and a Korean Neurological Association (KNA-20-Neuro-Frontier Fellowship Award Grant-7, 8).

Author Contributions

K.W.P. proposed the hypothesis. K.W.P., A.I., M.S.M., R.H., H.M., and M.J.M. designed the experiments. A.I. developed the data collection app for the UT dataset. F.S. collected and cleaned the UT dataset, and M.G. handled data collection and cleaning for the UBC dataset. A.I. processed videos, extracted features, designed the model, and conducted data analysis. The initial draft was created by K.W.P. and A.I. All authors reviewed the manuscript, provided feedback, and approved the final version for submission.

Competing interests

The authors declare no competing interests.

Djaldetti, R., Ziv, I. & Melamed, E. The mystery of motor asymmetry in Parkinson's disease. The Lancet Neurology 5, 796-802 (2006).
Bloem, B. R., Okun, M. S. & Klein, C. Parkinson's disease. The Lancet 397, 2284-2303 (2021).
Gómez-Esteban, J. C. et al. Factors influencing the symmetry of Parkinson's disease symptoms. Clinical neurology and neurosurgery 112, 302-305 (2010).
Baba, Y. et al. Phenotypic commonalities in familial and sporadic Parkinson disease. Archives of neurology 63, 579-583 (2006).
Yust-Katz, S., Tesler, D., Treves, T. A., Melamed, E. & Djaldetti, R. Handedness as a predictor of side of onset of Parkinson's disease. Parkinsonism & related disorders 14, 633-635 (2008).
Miller-Patterson, C. et al. Motor asymmetry over time in Parkinson’s disease. Journal of the neurological sciences 393, 14-17 (2018).
Imbalzano, G. et al. How resistant are levodopa‐resistant axial symptoms? Response of freezing, posture, and voice to increasing levodopa intestinal infusion rates in Parkinson disease. European Journal of Neurology 30, 96-106 (2023).
Maycas-Cepeda, T. et al. Hypomimia in Parkinson's disease: what is it telling us? Frontiers in Neurology 11, 603582 (2021).
Ricciardi, L. et al. Hypomimia in Parkinson’s disease: an axial sign responsive to levodopa. European Journal of Neurology 27, 2422-2429 (2020).
Guerra-Hiraldo, J. D. et al. Hemihypomimia in Parkinson’s disease: an under-recognized clinical sign? Journal of Neurology 270, 548-551 (2023).
Özekmekçi, S., Benbir, G., Özdogan, F. Y., Ertan, S. & Kiziltan, M. E. Hemihypomimia, a rare persistent sign in Parkinson’s disease: follow up of 11 subjects. Journal of neurology 254, 347-350 (2007).
Zingler, V. C., Strupp, M., Jahn, K. & Brandt, T. Hemihypomimia in Parkinson's disease. European neurology 53, 92-93 (2005).
Kurtis, M. et al. in Movement Disorders. S914-S914 (WILEY 111 RIVER ST, HOBOKEN 07030-5774, NJ USA).
Kazimierczak, N. et al. Skeletal facial asymmetry: reliability of manual and artificial intelligence-driven analysis. Dentomaxillofacial Radiology 53, 52-59 (2024).
Berssenbrügge, P. et al. 2D and 3D analysis methods of facial asymmetry in comparison. Journal of Cranio-Maxillofacial Surgery 42, e327-e334 (2014).
Szczapa, B. et al. in 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019). 1-5 (IEEE).
Hoehn, M. M. & Yahr, M. D. Parkinsonism: onset, progression, and mortality. Neurology 17, 427-427 (1967).
Caligiore, D. et al. Parkinson’s disease as a system-level disorder. npj Parkinson's Disease 2, 1-9 (2016).
Eidelberg, D. et al. The metabolic anatomy of Parkinson's disease: complementary [18F] fluorodeoxyglucose and [18F] fluorodopa positron emission tomographic studies. Movement disorders: official journal of the Movement Disorder Society 5, 203-213 (1990).
Yang, Y., Cheon, M. & Kwak, Y. T. 18F-FP-CIT positron emission tomography for correlating motor and cognitive symptoms of Parkinson's disease. Dementia and Neurocognitive Disorders 16, 57-63 (2017).
Kempster, P., Gibb, W., Stern, G. & Lees, A. Asymmetry of substantia nigra neuronal loss in Parkinson's disease and its relevance to the mechanism of levodopa related motor fluctuations. Journal of Neurology, Neurosurgery & Psychiatry 52, 72-76 (1989).
Karachi, C. et al. Cholinergic mesencephalic neurons are involved in gait and postural disorders in Parkinson disease. The Journal of clinical investigation 120, 2745-2754 (2010).
Bress, K. S. & Cascio, C. J. Sensorimotor regulation of facial expression–an untouched frontier. Neuroscience & Biobehavioral Reviews, 105684 (2024).
Pasquini, J. & Pavese, N. Striatal dopaminergic denervation and hypomimia in Parkinson's disease. European Journal of Neurology 28 (2021).
Schmidt, K. L., Liu, Y. & Cohn, J. F. The role of structural facial asymmetry in asymmetry of peak facial expressions. Laterality 11, 540-561 (2006).
Goetz, C. G. et al. Movement Disorder Society‐sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS‐UPDRS): scale presentation and clinimetric testing results. Movement disorders: official journal of the Movement Disorder Society 23, 2129-2170 (2008).
Qiang, J. et al. Review on facial-recognition-based applications in disease diagnosis. Bioengineering 9, 273 (2022).
Bandini, A. et al. Analysis of facial expressions in parkinson's disease through video-based automatic methods. Journal of neuroscience methods 281, 7-20 (2017).
Jin, B., Qu, Y., Zhang, L. & Gao, Z. Diagnosing ydisease through facial expression recognition: video analysis. Journal of medical Internet research 22, e18697 (2020).
Hou, X. et al. A markerless 2d video, facial feature recognition–based, artificial intelligence model to assist with screening for parkinson disease: development and usability study. Journal of Medical Internet Research 23, e29554 (2021).
Amsalam, A. S., Al-Naji, A., Daeef, A. Y. & Chahl, J. Automatic facial palsy, age and gender detection using a raspberry pi. BioMedInformatics 3, 455-466 (2023).
Brito, H. H. d. A. & Mordente, C. M. Facial asymmetry: virtual planning to optimize treatment predictability and aesthetic results. Dental Press Journal of Orthodontics 23, 80-89 (2018).
Bianchini, E. et al. The Story behind the Mask: A Narrative Review on Hypomimia in Parkinson’s Disease. Brain Sciences 14, 109 (2024).
Bradski, G. The OpenCV library. Dr. Dobb's Journal of Software Tools (2000).
Prados-Torreblanca, A., Buenaposada, J. M. & Baumela, L. Shape preserving facial landmarks with graph attention networks. arXiv preprint arXiv:2210.07233 (2022).
Baltrušaitis, T., Robinson, P. & Morency, L.-P. in 2016 IEEE winter conference on applications of computer vision (WACV). 1-10 (IEEE).
Lubba, C. H. et al. catch22: CAnonical Time-series CHaracteristics: Selected through highly comparative time-series analysis. Data Mining and Knowledge Discovery 33, 1821-1852 (2019).
Chen, T. & Guestrin, C. in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining. 785-794.

No competing interests reported.

Quantitative Assessment of Facial Expression Asymmetry in Parkinson’s Disease

Status:

Version 1

Abstract

Figures

Introduction

Results