A Machine Learning System for Quantifying Facial Asymmetry in Bell’s Palsy across Underrepresented Phenotypes

Facial paralysis, often known as Bell's palsy, is a sudden condition where the muscles on one side of the face weaken or stop working entirely. It can leave a person with a drooping mouth, a closed eye that won't open, and a forehead that refuses to wrinkle. For doctors, judging how severe this condition is and tracking whether a patient is getting better usually relies on a trained eye. A specialist looks at the face and assigns a score based on how much movement remains. This method works well when a specialist is present, but in many parts of the world, particularly across South Asia, such experts are rare. Most patients are first seen by general practitioners who may not have the same level of training. To make matters more difficult, human judgment can vary from one doctor to another, and small improvements in a patient's condition are hard to spot when they are measured by a subjective scale rather than a precise number.

Researchers have long hoped that computers could solve this problem by taking a photograph of a patient's face and measuring the difference between the left and right sides with perfect consistency. The idea is to build a tool that works on ordinary hardware, like a standard laptop or tablet, so it can be used in clinics without expensive equipment. However, a major concern has been whether these computer tools work equally well for people of all backgrounds. Many facial recognition systems have historically been trained on datasets that mostly feature people with lighter skin tones, leading to fears that they might fail or give inaccurate readings for people with darker skin or different facial features. If a tool meant to help everyone fails silently on specific groups of people, it could do more harm than good.

A team of researchers from King Edward Medical University set out to test a new computer system designed specifically for this task, focusing on whether it could reliably measure facial asymmetry across a wide variety of human faces. They did not test the system on patients with Bell's palsy yet; instead, they used a large collection of photographs of healthy people from seven different demographic groups to see how the tool behaved when there was no disease present. Their goal was to ensure the system was stable and fair before it was ever used on a sick patient. They built a pipeline that first finds the face in a photo, then maps hundreds of points across the skin to create a digital outline, and finally draws an imaginary line down the center of the face to measure how far each point sits from that center.

The system they built uses a pre-trained software component that can instantly find 478 specific points on a face from a single image. Once it has these points, it calculates a score that represents how uneven the face is. A score of zero would mean a perfectly symmetrical face, while a higher score indicates more asymmetry. The researchers tested this on over 10,000 images from a public dataset called FairFace, which was carefully constructed to include balanced numbers of people from East Asian, Southeast Asian, Indian, Black, Latino Hispanic, White, and Middle Eastern backgrounds. They wanted to see if the computer could find the face in the photo and if the resulting asymmetry score was consistent across all these different groups.

The results showed that the system is largely effective but has some specific quirks. It successfully found the face and calculated a score in about 90 percent of the images. However, the rate at which it failed to find a face was not the same for everyone. Contrary to the common expectation that such tools struggle more with non-Western faces, this system actually failed more often on White and Middle Eastern faces than on Southeast Asian faces. The failure rate was lowest for Southeast Asian faces at roughly 6 percent and highest for White faces at nearly 13 percent. The researchers noted that this reversal likely stems from differences in the source of the photos, such as variations in lighting, the presence of beards or glasses, or the angle of the head, rather than a fundamental flaw in how the computer sees different skin tones. This finding suggests that before deploying such a tool in a new region, it is essential to test it directly on the local population rather than assuming it will behave the same way as it did in a different dataset.

When the system did successfully measure the faces of healthy people, the results were remarkably consistent. The average asymmetry score for a healthy face was very low, hovering around 0.187. While statistical tests showed tiny differences between the groups, these differences were so small that they would not matter in a real-world clinical setting. The system proved that it could measure a healthy face with the same baseline precision regardless of the person's demographic background. One specific part of the measurement, however, stood out as unreliable: the area around the inner corner of the eye. Because this part of the face is so close to the center line, the math used to calculate the score became unstable, producing artificially high numbers that did not reflect actual facial structure. The researchers identified this as a mathematical artifact rather than a real biological finding, noting that the tool works best when looking at features further from the center, like the mouth or the outer eye.

To make the tool useful for doctors, the researchers also established a way to flag images that might need attention. They calculated a threshold score for each group, set so that only the top 5 percent of healthy faces would be flagged as potentially abnormal. These thresholds were very similar across all groups, with only a small variation of about 8.8 percent between the lowest and highest. This means that a single standard score could likely be used for most people, though the system might need a slightly adjusted rule for Southeast Asian faces to avoid raising too many false alarms. The researchers were careful to state that this system currently only controls for false alarms on healthy people; it has not yet been tested to see if it can correctly identify actual cases of Bell's palsy. That is the next step.

The study concludes that this lightweight computer system is ready to move from a theoretical test to a real-world trial, provided that doctors understand its limitations. It is a tool that can run on ordinary hardware and give consistent measurements across diverse populations, but it is not yet a diagnostic device. The researchers emphasize that the system's ability to detect the disease itself remains unproven and must be validated in a clinical study with actual patients. Until then, the tool serves as a reliable way to measure facial symmetry in healthy individuals, offering a promising path toward more consistent and accessible care for people with facial paralysis in resource-limited settings. The work highlights that before a machine can help a doctor, it must first prove it can see the world clearly, without bias or error, across the full spectrum of human faces.