AI-supported exercise improved walking capacity in a hypertension trial

Adults with hypertension walked farther after an eight-week AI-supported exercise programme than those receiving conventional guidance. The 62-person randomised Guangzhou trial, published on 13 January 2026, found a baseline-adjusted 62.77-metre advantage on a six-minute walk test (95% confidence interval 26.33–99.22).[1]

Both groups received prescriptions from the same exercise library. The app used OpenPose to provide movement feedback, alongside reminders and weekly professional consultation; controls had no active exercise supervision. The trial cannot isolate automated feedback from professional follow-up.[1]

The walking test measures capacity under set conditions. A person walks as far as possible along a flat corridor, choosing their own pace, and the distance is recorded. The European Respiratory Society and American Thoracic Society’s technical standard stresses consistent instructions, layout and encouragement. These details make the comparison meaningful: the result describes performance on a defined task, rather than the number of steps someone happened to take during an ordinary day.[2]

The open-label analysis retained 31 people per arm and imputed four missing primary follow-ups. Participants needed smartphones and exercise screening.[1]

The computer-vision mechanism turns a camera image into a representation of posture. In the original work behind OpenPose, one branch of the network estimates the locations of body parts; another estimates the connections between them. This allows it to assemble a person’s pose from points such as elbows and wrists. Linking the right parts matters: finding several joints is insufficient if the software cannot tell which belong together.[3]

The coordinates let software estimate an angle from a shoulder, elbow and wrist, then follow it across frames. This supplies the movement information needed for feedback on a person’s own posture.[3]

Report an error or suggest a correction