Smartphone-based Facial Palsy Detection using Landmark Asymmetry and Lightweight SVM Classification
Authors: Holi, P., Dubey, V., Shearer, C., Thavanesan, K.
Conference: IEEE-EMBS International Conference on Body Sensor Networks 2026
Dates: 10/10/2026
Publication Date: 10/10/2026
Abstract:Facial Palsy is an important clinical indicator associated with neurological conditions like Stroke and Bell’s palsy. Early and accurate detection is essential for timely intervention and improved patient outcomes. Traditional assessment methods often rely on subjective clinical observation. This paper proposes a lightweight smartphone-based facial palsy screening system using facial landmark asymmetry analysis and Support Vector Machine (SVM) classification. The system integrates Google ML Kit for landmark detection, geometric symmetry analysis, and TensorFlow Lite (TFLite) to enable real-time processing on the smartphone device. It extracts clinically meaningful asymmetrical features from facial landmarks to assess one-sided facial muscle imbalance. The developed smartphone application was evaluated using a public dataset containing both healthy subjects and those with facial palsy under various lighting conditions, facial expressions, and head positions. Five-fold cross-validation achieved 95.2 ± 1.1% accuracy. Independent testing yielded 95.88% accuracy, 91.0% specificity, 99.07% sensitivity, and 94.90% F1 score. Comparisons with Random Forest, K-Nearest Neighbour (KNN), CNN, ResNet50, and EfficientNet-Lite0 models show that the proposed method demonstrated strong performance while significantly lowering computational demands, processing time, and inference delay. Additionally, it does not require large datasets to deliver effective results, relying instead on geometric analysis and interpretable ML techniques. Overall, our approach aims to provide a portable, accessible, and cost-effective solution for initial patient assessment and early detection of facial palsy for neurological screening.
Source: Manual