Single-camera pose estimation for quantitative assessment of balance deficits: a proof-of-concept using graded acute alcohol intake

Balance is the quiet, constant negotiation between the brain and the body. It relies on a continuous loop where the eyes, the inner ear, and the sensors in the muscles and joints send signals to the brain, which then makes tiny, invisible adjustments to keep us upright. When this loop is disrupted, as it is in conditions like stroke or Parkinson's disease, the ability to stand still or walk safely can vanish. For decades, doctors have relied on a standard checklist called the Berg Balance Scale to measure this ability. A clinician watches a patient perform a series of tasks, such as standing on one foot or turning around, and assigns a score from zero to four based on how well the task was completed. While this method is widely used, it has a fundamental flaw: it is a blunt instrument. It records a final judgment of success or failure, but it cannot see the subtle tremors, the slight sways, or the micro-adjustments that happen before a person actually falls. Worse, for people who are already quite steady, the scale hits a ceiling. If a patient can stand safely for thirty seconds, they get the highest possible score, even if their balance is slightly worse than it was yesterday. This makes it nearly impossible to track small improvements or declines in people who are functioning well, which is exactly when precise measurement is most needed for rehabilitation.

To solve this problem, researchers at ETH Zurich and the Cantonal Hospital Baden set out to see if a simple video camera could detect changes in balance that a human observer misses. They needed a way to create a known, controlled change in balance that could be measured precisely. They chose to use alcohol. By having healthy volunteers drink specific amounts of alcohol, the researchers could induce a temporary, graded impairment of the cerebellum and vestibular system—the very parts of the brain that control balance. This allowed them to test a new digital method against the standard clinical scale under conditions where the "true" level of impairment was known. The team recorded twenty-four healthy adults performing six different standing tasks, ranging from standing on two feet to standing on one foot with eyes closed. They did this at three different levels of alcohol in the blood: zero, a low dose of 0.6 per mille, and a high dose of 1.0 per mille. Every session was filmed with a single, ordinary consumer camera, the kind found in a smartphone, positioned in front of the participant.

The researchers then used advanced computer software to analyze the video. Instead of asking a human to guess how steady a person looked, the software tracked the movement of the person's body frame by frame, creating a digital skeleton of their posture. From this video, the computer calculated 167 different measurements of movement, such as how fast the body swayed side-to-side, how much the trunk moved, and whether the person took a step to regain balance. The team selected the most sensitive of these measurements and combined them into a single digital score. They then compared this digital score to the score given by a trained human rater watching the same video, using a simplified version of the standard balance scale. The human rater assigned a score from zero to four based on how long the person could hold the position without falling or needing help.

The results showed a clear difference in what the two methods could see. The human rater, using the standard scale, gave the maximum score of four in 77 percent of all the trials. This means that for most of the time, even as the participants drank more alcohol and their balance worsened, the human observer saw no change at all. The scale had hit its limit; it could not distinguish between a person who was perfectly steady and a person who was slightly unsteady but still standing. In contrast, the digital score derived from the single camera continued to register the decline. The computer-based measure was able to correctly rank the three levels of alcohol intake for the majority of participants, achieving a mean accuracy of 86 percent, whereas the human rating only did this correctly for about 59 percent of the time. The digital score was significantly better at detecting the difference between the sober state and the low-dose state, and even more so between the low and high doses.

Perhaps the most revealing finding occurred in the group of participants who received the maximum score of four from the human rater at both the sober and the high-alcohol levels. To the human eye, these people appeared unchanged. However, the digital score showed a significant, steady increase in body sway as the alcohol level rose. The computer detected a graded deterioration in balance that the human observer completely missed. This demonstrated that a maximum score on the standard scale does not mean a person's balance is perfect; it only means their performance fell within the top category of the scale. The digital method, by measuring the actual movement in physical units rather than a subjective judgment, could see the impairment that was hidden by the ceiling of the traditional test.

The study also examined how reliable the human raters were. When two different experts watched the same videos, they agreed on the exact score 83 percent of the time. However, this high agreement was largely because they both gave the maximum score to the same people. When the researchers looked only at the trials where the score was not the maximum, the agreement between the two experts dropped to just 30 percent. This suggests that when a human has to make a real decision about a borderline case, their judgment varies significantly. The digital score, being based on a consistent mathematical calculation of movement, did not suffer from this variability.

The researchers concluded that a single, ordinary camera can capture fine details of balance that a standard clinical rating cannot. The digital score was able to track the effects of alcohol dose by dose, even in people who were rated as having no balance problems at all. This does not mean the human observer is useless, but it does show that the traditional scale has a blind spot for people who are functioning well. The digital approach offers a way to measure balance continuously and objectively, without the need for expensive equipment, special markers on the body, or a laboratory setting. It suggests that in the future, a simple video could provide a much more sensitive tool for monitoring recovery or decline in patients, catching changes that would otherwise go unnoticed until they became too large to ignore. The study proves that the camera sees what the eye cannot, turning a simple recording into a precise measure of stability.