An AI tool that ‘fine-tunes’ the assessment of patients with facial palsy has been hailed as a significant step in incorporating the technology into clinical practice.
An experimental study led by Takeichiro Kimura from Kyorin University in Mitaka, Tokyo, suggests that research offers valuable insights into using AI when treating patients with facial palsy who experience paralysis or partial loss of facial movement caused by nerve injury from tumours, surgery, trauma, or other factors.
A thorough assessment is essential for evaluating treatment options, like nerve transfer surgery, but it also presents significant challenges.
Various subjective scoring systems have been developed, but they are subject to variability.
Objective assessments have been described but remain impractical for routine clinical use. Machine learning and AI models offer a potential solution for regular, objective evaluation of facial palsy.
Dr Kimura and colleagues evaluated a previous AI-based facial recognition model, called 3D-FAN, in patients with facial palsy. This system was trained to identify 68 facial key points, including eyebrows, eyelids, nose, mouth, and facial contours.
When used on clinical videos, 3D-FAN, which was trained on images of people with normal facial movement, proved insufficient for evaluating facial palsy.
The system frequently failed to identify facial asymmetry, including when patients were instructed to smile, and did not recognise when eyes were closed.
AI tool shows promise for objective ratings of facial palsy severity.
Dr Kimura and colleagues attempted to fine-tune the model using machine learning, based on 1,181 images extracted from clinical videos of 196 patients with facial palsy.
During this process, facial landmarks were manually repositioned to their correct locations, with an effort to minimise variability.
Training sessions were repeated until no further improvements in accuracy were observed.
Dr Kimura and colleagues said: ‘After machine learning, we observed both qualitative and quantitative enhancements in the AI’s ability to detect facial keypoints.’
The improved model had notably lower error rates, with enhanced keypoint detection across all facial regions, including the eyelids and mouth, which are crucial areas that exhibit asymmetry in facial palsy.
The authors believe their fine-tuning method, which includes manual correction of landmarks on a limited number of images, ‘has potential for wider application in developing AI-assisted models for other relatively rare disorders’.
Pending further validation, the researchers plan to make their AI model freely accessible to other researchers and clinicians.
Dr Kimura and co-authors concluded: ‘Considering our software as one of the promising solutions for objective assessment of facial palsy, we are now conducting a multidisciplinary analysis of its efficacy.’
By providing an objective score, the AI tool could facilitate more precise grading of facial palsy severity and serve as a quantitative measure for treatment outcomes.
The study was published in the June issue of Plastic and Reconstructive Surgery, the official medical journal of the American Society of Plastic Surgeons (ASPS).
