Abstract
Parkinson's disease (PD) severity diagnosis is crucial for early detecting potential patients and adopting tailored interventions. Diagnosing PD based on facial expression is grounded in PD patients' "masked face" symptom and gains growing interest recently for its convenience and affordability. However, current facial expression-based approaches often rely on single type of expression which can lead to misdiagnosis, and ignore the class imbalance across different PD stages which degrades the prediction performance. Moreover, most existing methods focus on binary classification (i.e., PD / non-PD) rather than diagnosing the severity of PD. To address these issues, we propose a new facial expression-based method for PD severity diagnosis which integrates multiple facial expression features through attention-based feature fusion. Moreover, we mitigate the class imbalance problem via an adaptive class balancing strategy which dynamically adjusts the contribution of training samples based on their class distribution and classification difficulty. Experimental results demonstrate the promising performance of the proposed method for PD severity diagnosis, as well as the efficacy of attention-based feature fusion and adaptive class balancing.
Introduction
Parkinson’s Disease (PD) is often associated with "hypomimia," or "masked face"—a clinical symptom where facial muscle movements are reduced, making the face appear expressionless. While researchers have previously used this as a biomarker for binary diagnosis (PD vs. non-PD), this paper tackles two major hurdles that have limited the clinical utility of facial analysis: the reliance on single-expression analysis and the lack of severity-based diagnosis.
Most existing methods rely on a single type of expression (like a smile). However, many patients in the early stages of PD can still produce certain expressions, leading to frequent misdiagnosis. Furthermore, datasets for PD are often "imbalanced," meaning there are far more samples of healthy individuals or early-stage patients than those in the mid-to-late stages, which biases the AI's learning process.
Proposed Methodology
The proposed pipeline moves beyond simple image classification through two novel components designed to capture a holistic "motor signature" of the face:
- Attention-based Feature Fusion (AFF): Instead of looking at one photo, the model takes six images of a subject attempting different basic emotions (anger, disgust, fear, happiness, sadness, and surprise). A ResNet-18 backbone extracts features from each. These features are then fused using an attention mechanism that automatically learns which emotional expressions are most indicative of PD severity for a specific individual.
- Adaptive Class Balancing (ACB): To solve the data imbalance problem, a custom loss function ($L_{adapt}$) was designed. This function gives more weight to minority classes (like late-stage PD) and uses a focal loss term to force the model to focus on "hard" samples that it previously misclassified.
Experimental Results
The study utilized the PDFE Dataset, containing images from 145 PD patients labeled by severity, alongside public benchmarks for non-PD control samples. The proposed method was tested against state-of-the-art (SOTA) models typically used for image recognition, such as MobileNetV4 and Swin-Transformer.
As shown in the results table, the model achieved an AUC of 0.8816 and an Accuracy of 83.54%, significantly outperforming standard architectures. Ablation studies (Ours w/o AFF and Ours w/o ACB) confirmed that both the attention mechanism and the adaptive balancing strategy are essential for peak performance, as removing either component led to a drop in accuracy and G-Mean.
Conclusion and Future Work
By integrating multiple emotional expressions and addressing class imbalance, this method provides a more robust and clinically relevant tool for diagnosing PD severity. While the results are promising, future work will explore dynamic video sequences to capture the latency of facial muscle activation and seek to expand the cohort size to further validate the tool for clinical use.
Summary updated