49. Development of a Novel Multi-feature Machine Learning Model for Unilateral Facial Paralysis

PURPOSE: Accurate diagnosis of unilateral facial paralysis is crucial for effective treatment. This study proposes a novel machine learning approach using the MediaPipe model to differentiate normal faces from paralyzed faces based on facial asymmetry indices. METHODS: A dataset of 111 healthy and 104 unilateral paralyzed smiling faces was analyzed. Six asymmetry indices were developed: visibility of teeth area, nasolabial fold depth, commissure ratios, smile index ratio, and visible teeth area. Logistic regression and multi-feature classifiers were applied. RESULTS: Single feature classifiers achieved accuracies of 65-78%. The best multi-feature classifier, combining smile index ratio, eye area difference, and teeth area difference, achieved 82% accuracy. Logistic regression on these three features further improved accuracy to 86%. CONCLUSION: Our proposed machine learning approach demonstrates high accuracy in diagnosing unilateral facial paralysis. Artificial intelligence machine learning models can provide an objective tool for physicians to assess facial paralysis and monitor treatment outcomes.