Dynamic Facial Analysis for Predicting Facial Palsy Outcomes: Comparing Landmark Detection Models and Integrating Ordinal Regression.

UNLABELLED: This study aims to enhance the prediction and video-based assessment of facial nerve (FN) recovery in facial palsy patients through incorporating modern landmark detection models and regression techniques. Our goal is to determine if these methods offer significant improvements to our previously reported predictive framework over conventional approaches.

METHODS

We extend our previous methodology by comparing state-of-the-art facial landmark detection models, such as ones that use deep learning, with Dlib. These models are evaluated based on their accuracy, computational cost, and impact on clinical score predictions. Additionally, we replace our previous least-squares linear regression model with ordinal regression to predict House-Brackmann (HB) scores, leveraging Wasserstein and Mahalanobis distances to better capture the ordered nature of the HB grading system.

RESULTS

Dlib offered the best balance of computational efficiency and clinical accuracy, while other higher-resolution models did not improve performance in predicting clinical scores. Ordinal regression significantly outperformed naive linear regression, demonstrating better interpretability, improved accuracy, and reduced mean absolute error by properly accounting for the ordinal structure of the HB scale.

SIGNIFICANCE: This study extends our previous work by incorporating modern landmark detection techniques and a more clinically appropriate predictive model for FN assessment. By bridging the gap between computational models and real-world clinical applications, this framework enhances the precision of facial palsy monitoring, offering a more robust tool for surgical decision-making and longitudinal patient assessment.