Several prognostic models for neck pain have been developed. Why aren’t these ready for physiotherapists to use?

Studies of prognostic models comprise three consecutive stages: model development with internal validation, external validation in new settings and assessment of a model’s clinical impact.

Before a developed prognostic model can be considered for use in clinical practice, it is essential to assess at least its external validity to understand the model’s transportability in a setting comparable to its intended use.

However, most models don’t reach the external validation phase. Methodological shortcomings are common in studies of prognostic models.

Too many candidate predictors relative to the number of outcome events are analysed, predictor selection is based purely on statistical significance, continuous predictors are categorised or limitations exist in the measurement of predictors and outcomes.

Such shortcomings often lead to overfitted, overoptimistic or unstable/biased models that generalise poorly to other clinical settings and patients, underlining the need for external validation.

What were the three existing models that you submitted to external validation?

We submitted a model predicting neck pain recovery to external validation in terms of 1) disability, 2) pain intensity and 3) perceived improvement.

Roel Wingbermühle.

These three outcomes were assessed at six and 12 weeks and at the end of the treatment period. The disability model was especially promising since it showed acceptable performance at internal validation.

What was the cohort in which you validated them?

We used PRONEPA, a prospective cohort study that ran from November 2020 to April 2021 and included a convenience sample of 586 participants with neck pain for less than 12 weeks, recruited by 102 physiotherapists who were graduating from a Master of Science program in manual therapy at SOMT University of Physiotherapy in Amersfoort, the Netherlands.

Patients received usual care physiotherapy interventions in accordance with Dutch clinical practice guidelines, which included information on the benign nature of the condition, advice to stay active, neck muscle strengthening exercises and cervical spine mobilisations.

What did you find? Were any of the models suitable for use after validation?

The disability model for prediction of neck pain recovery at six weeks showed acceptable discriminative performance (c-statistic equal to 0.73 (95 per cent CI 0.69 to 0.77)) and was reasonably well calibrated after intercept recalibration.

The pain intensity model and perceived improvement model did not reach acceptable levels of performance measures (ie, discrimination and calibration).

With reasonable confidence, we advise the use of the disability model in clinical practice at intake for the prognosis of patients with acute or subacute neck pain to assist in clinical decisions concerning recovery from neck pain-related disability at six weeks.

The disability model’s regression formula is complex. Is there anything you can do about this?

A model’s regression formula usually is complex. Therefore, we present it as a web-based risk calculator (somt.nl/en/research/ neck-pain-recovery-prediction-calculator).

On the website, the physiotherapist fills in the individual patient’s variables: age, sleep problems yes/no, neck pain duration within 6–12 weeks yes/ no, sum score of the Neck Disability Index and answers to four questions from the Fear-Avoidance Beliefs Questionnaire (physical activity subscale).

The calculator provides the probability of non- recovery from neck pain disability for this patient after six weeks.

What do you think research in this field should address next?

The relatively low predictive performance of neck pain models indicates that predictors are still missing and that the quest for additional predictors needs to continue.

Several problems in neck pain modelling can be alleviated if large studies specifically designed for prognostic model research use baseline standard measurement sets that are tuned to cover a wide array of biological,physical and psychosocial measures.

New methods for analysing complex networks of interacting variables, such as machine learning techniques, may be a promising way to account for the complex nature of neck pain.

>> Roel Wingbermühle is a physiotherapist who combines his clinical work as a manual therapist with educational work and scientific research. This external validation study was the final study for his doctorate and he obtained his PhD in December 2022.