Beyond One Accuracy Score: A Multidimensional View of Cervical Movement Control

Cervical movement control is unlikely to be captured fully by one accuracy score. A 2026 study using the Butterfly test found that combinations of directional-accuracy measures across several task difficulties distinguished a chronic neck-pain group from asymptomatic controls more consistently than most isolated measures. For clinicians, the useful message is not that a machine-learning model can diagnose neck pain. It is that how a person misses a target—and what happens as the task becomes harder—may matter alongside the average size of the error.

That distinction is important. Two people can finish a tracking task with a similar mean deviation while producing very different movement patterns. One may spend much of the trial just inside the target path, another may alternate between overshooting and undershooting, and a third may be accurate on an easy path but lose control as direction and speed become less predictable.

What did the study ask?

Ziva Majcen Rosker and Jernej Rosker investigated whether measures from a cervical movement-control test could classify people with chronic idiopathic neck pain and asymptomatic controls. They also asked whether combining several measures and difficulty levels gave a more balanced result than considering one measure at a time.

The study included 65 people with chronic idiopathic neck pain and 50 age- and sex-matched asymptomatic participants. People with previous head or neck trauma, arm or shoulder pain, neurological or vestibular disorders, type II diabetes or psychiatric disorders were excluded. This created a reasonably well-defined comparison, but one that is narrower than the mixed presentations seen in many MSK, concussion and vestibular clinics.

How was movement control measured?

Participants completed the Butterfly test using the NeckCare NeckSmart system. A head-mounted inertial sensor tracked head movement while the participant followed an unpredictable target on a computer screen.

The task used three predefined paths—easy, medium and difficult—presented in random order and repeated three times. The analysis included five performance features:

• Amplitude accuracy: the average distance between the participant's trace and the target.
• Time on target: the proportion of the task spent within the target area.
• Undershoot: the proportion of movement falling inside the target path.
• Overshoot: the proportion extending outside it.
• Smoothness: a measure of how fluently the movement was produced.

The authors grouped time on target, undershoot and overshoot as “directional accuracy”. They then compared several machine-learning approaches using individual features and combinations of features across the three difficulty levels. The models were assessed with five-fold cross-validation within the same dataset.

What did the researchers find?

Among the individual measures, amplitude accuracy and measures related to time on target or undershoot showed the strongest classification performance. The most balanced results came from combining directional-accuracy measures across all difficulty levels, or from combining the full set of parameters across all difficulty levels.

This is a reported difference in how well study groups could be distinguished statistically. It should not be translated into a clinical diagnostic test. The paper did not provide an independently validated model for deciding whether an individual patient's neck pain is present, absent or caused by a particular movement-control impairment.

Why one mean error can hide useful information

A single summary value is attractive because it is quick to record and easy to compare. The trade-off is compression: different movement strategies can end in the same average.

Consider a visual-tracking task. Mean deviation tells us the overall distance from the target. It does not reveal whether the person consistently fell short, repeatedly passed beyond the path, stayed close to the target for most of the trial but made one large correction, or moved with frequent small oscillations. Those patterns may create different questions for clinical observation even when the final mean is similar.

Difficulty adds another dimension. Performance on a predictable or slow task may not describe control when the path becomes more complex. A graded test can therefore be informative without implying that “harder is better”. The task should remain safe, relevant and standardised, and its interpretation should stay within the evidence available for that method.

Practical takeaways for clinicians

The study supports several cautious principles for assessment and review:

1. Define the feature you are observing. Accuracy, direction of error, time on target and smoothness are related, but they are not interchangeable.
2. Record task conditions. Target size, distance, movement range, speed, predictability, instructions, repetitions and symptoms can all affect performance.
3. Use graded difficulty deliberately. An easier condition can establish the task; a more demanding condition may reveal a different control strategy. Compare like with like at reassessment.
4. Look beyond the final score. Note repeated undershooting, overshooting, corrective movements, pauses, substitution and symptom response where clinically relevant.
5. Do not turn group discrimination into diagnosis. A statistical model that separates two research groups is not automatically a validated decision tool for an individual patient.
6. Keep measurement separate from treatment claims. This study did not test rehabilitation, clinical outcomes or change over time.

Where visual feedback fits

Head-mounted visual feedback can make a head-movement path visible to a patient and observable to a clinician. In a clinician-selected tracing task, a laser may help reveal the direction and pattern of error rather than only the final position. That makes the multidimensional measurement principle relevant to HeadX's clinical audience.

However, the paper used the NeckCare NeckSmart IMU and software. It did not study HeadX Kross or HeadX Duo, and a laser-tracing task does not reproduce the study's sensor data, algorithms or classification model. Any HeadX exercise should therefore be framed as a separate clinician-led activity, not as an implementation or validation of the Butterfly-test research.

Important limitations

The study was cross-sectional, so it cannot show whether altered movement control causes neck pain, results from it, predicts future symptoms or changes with treatment. Its machine-learning models were tested with internal cross-validation but not in an independent external cohort. The sample was modest for multivariable modelling and excluded several clinically important presentations, including previous head or neck trauma and vestibular or neurological disorders.

The study also did not establish diagnostic cut-offs, minimal detectable change, clinically important change, test-retest performance for longitudinal monitoring or improved patient outcomes. The observation that some asymptomatic people may show less controlled movement cannot be treated as evidence that they will develop neck pain or benefit from preventive intervention.

In summary

The paper's strongest contribution is conceptual: cervical movement control can be described across several features and levels of challenge. Average accuracy still has value, but it may not tell the whole story. For clinicians, a fuller assessment can mean observing where the trace goes, how long it stays on target, how smoothly it moves and how performance changes when the task becomes more demanding—while resisting the temptation to convert a promising research classifier into a diagnosis or product claim.

Suggested internal links

• The Cervical Joint Position Error Test: a clinical guide
• Cervical range of motion normal values
• HeadX exercise programme
• Chapter B: Proprioceptive Accuracy & Repositioning
• HeadX Kross

References

1. Majcen Rosker Z, Rosker J. Multidimensional machine learning approach for classifying patients with neck pain based on movement control test. Brazilian Journal of Physical Therapy. 2026;30(4):101606. PubMed | Full text
2. AlDahas A, Devecchi V, Deane JA, Falla D. Measurement properties of cervical joint position error in people with and without chronic neck pain. PLOS ONE. 2023;18(10):e0292798.

This article is written for qualified clinical professionals. It is educational and does not constitute medical advice. The featured study did not evaluate HeadX or establish the clinical performance of a HeadX product.