Dynamic and Objective Quantification of Facial Nerve Palsy Using Smartphone-Based 3D Facial Geometry

To test whether smartphone-captured high-resolution 3D facial geometry can objectively quantify the severity of facial nerve palsy (FNP).

We enrolled 99 patients with peripheral acute unilateral FNP (207 visits) and 12 healthy volunteers (July 2021-February 2024). The facial dynamics were captured using a smartphone with a 3D front-facing camera, yielding a facial mesh comprising 1220 vertices. Recordings were obtained for multiple facial tasks. Clinical benchmarks included the House-Brackmann (HB) grading and electrophysiology. Features of displacement, velocity, acceleration, and asymmetry were extracted. Random forest regressors were trained for signed continuous HB (side-aware) score prediction and for region-specific prediction. Patient-wise cross-validation was used. Regressor performance was evaluated using Spearman's ρ, mean absolute error, R², and area under the curve for HB ≥ III.

The continuous HB regression model achieved a mean absolute error of 0.718, an R² of 0.851, and a Spearman's ρ of 0.935. Predictions for healthy controls were tightly clustered. The area under the receiver operating characteristic curve for screening for HB ≥ III was 0.914 ± 0.052. Region-specific analyses showed the best performance for eyelid closure and teeth exposure. Regression using integrated electromyography demonstrated moderate correlations, while electroneurography showed weak correlations.

Smartphone-captured high-frame-rate 3D facial geometry enabled accurate side-aware continuous HB estimation and robust screening for clinically significant palsy. This method can detect subtle residual asymmetry beyond human grading, is robust for healthy controls, and may facilitate early detection, standardized monitoring, and at-home rehabilitation of FNP.