Multidimensional machine learning approach for classifying patients with neck pain based on movement control test.

The sample was small, 65 patients and 50 controls from a single setup, and the authors say the models may still be overfitted and call for larger studies to confirm the results. The Butterfly test measures have no published reliability data, so part of the difference between people may be measurement noise rather than real impairment. The study protocol was not registered. The authors say the cut-off values from their sample cannot be used to guide clinicians, because patients with neck pain are a heterogeneous group and their values differ from those reported in other studies. The control group was not checked for ergonomic or occupational risk factors such as occupation, working hours and education, which may be why some of them were misclassified as patients. Only pain intensity was recorded; other symptoms such as central sensitisation were not measured, so the link between symptoms and the extent of movement control impairment could not be explored. The models were never tested on a separate group of people, so how they would perform on new patients is unknown.

The authors declare they have no competing interests. The Butterfly test was delivered with a commercial system, NeckCare Holding's NeckSmart software with a head-mounted inertial sensor.

Do not treat this as proof that the Butterfly test is ready to diagnose neck pain in the clinic. Some people without neck pain were classified as patients, the authors report that single-measure models were too over- or under-confident to be suitable for clinical use, they warn that their own cut-off values cannot be generalised to guide clinicians, and the whole analysis rested on 65 patients and 50 controls.