Design
This is a prospective cohort study with a 6- and 12-month follow-up. The Medical Ethics Committee of the Elisabeth Hospital in Tilburg, The Netherlands, approved the study. All participants provided written informed consent prior to participating.
Participants
Participants were recruited between July 2013 and October 2014. Consecutive patients with cervical radiculopathy who were referred to a multidisciplinary clinic in The Netherlands by their general practitioner or medical specialist were eligible for participation. All participants underwent MRI scanning before entering the study. A neurosurgeon with extensive (i.e., > 10 years) clinical experience in managing patients with cervical radiculopathy reached the diagnosis of cervical radiculopathy if clinical findings from the history and physical examination (e.g., pain, numbness, paresthesia, muscle strength, and reflex changes) corresponded with nerve root compression observed on MRI. Inclusion criteria for this study were: diagnosis of cervical radiculopathy due to disk herniation, stenosis or a combination, at least 18 years of age, referred for conservative management and adequate understanding of the Dutch language to complete the questionnaires. Patients were excluded in case of known serious pathology (such as malignancies, fractures, (rheumatoid) arthritis, infections or myelopathy), multiple sclerosis, diabetes mellitus, polyneuropathy, complex regional pain syndrome or a history of spinal surgery.
Procedure
At baseline, patients provided information regarding demographics and potential prognostic factors via electronic questionnaires. The neurosurgeon performed a clinical neurological examination. After 6 and 12 months, patients completed a digital survey of questions regarding the current level of recovery (Global Perceived Effect scale [9]); questions regarding their level of symptoms (including Numeric Pain Rating Scales for neck pain, arm pain and disability [10]); sick leave due to the cervical radiculopathy (duration in weeks); treatment received (i.e., physical therapy, manual therapy, injections, medication, other) and medication use (type and amount). A copy of the digital survey is provided in Appendix 1. Participants who did not respond to the electronic questionnaire, received a reminder after 1 and 2 weeks, followed by a final reminder via a telephone call.
Outcomes
The course of cervical radiculopathy was described in terms of perceived recovery, neck and arm pain intensity and perceived disability at 6 and 12 months. Additionally, we determined the proportion of participants with a high pain intensity at 6 and 12 months, i.e., a score of 7 or higher on an 11-point Numeric Rating Scale (NRS) [10, 11].
The primary outcome measure for the prognosis was the perceived recovery at 12 months, measured on a 7-point Global Perceived Effect (GPE) scale [9]. Patients were considered recovered if they scored ‘completely recovered’ or ‘much improved’ [9]. Secondary outcome measures were neck pain intensity and disability level at 12 months. Patients were considered recovered if they scored ≤ 2 for neck pain intensity and disability on an 11-point NRS, ranging from 0 to 10 [12].
Potential predictors
We determined which factors to include in the multivariable analyses for each outcome measure separately [13, 14]. Because there is a lack of knowledge about prognostic factors for cervical radiculopathy, we included prognostic factors for non-specific neck pain, such as duration of symptoms (weeks), previous episodes of neck pain (yes/no), pain intensity (0–10) and presence of low back pain (yes/no) [15, 16]. Additionally, because we were interested in physical factors that could be influenced by conservative management, the following potential prognostic factors were included: active range of motion of the neck (measured with a Cervical Range or Motion device (CROM; Performance Attainment Associates, Lindstrom, MN, USA)) [17]; deep neck flexor endurance (measured with a clinical muscle endurance test as described by Harris et al. [18]); the level of disability (measured with an 11-point NRS, ranging from 0 (no disability) to 10 (total disability)) and the presence of neuropathic pain (measured with the Dutch language version of the PainDETECT Screening questionnaire) [19]. The factors needed to be easily obtainable and reliable to measure in clinical practice, to ensure that the factors can be widely used in clinical practice. Table 1 provides an overview of the selected potential predictors per outcome measure.
Table 1 Overview of predictors included in the multivariable logistic regression analyses per outcome Statistical analysis
Missing values
We performed several missing value analyses: First, we performed Little’s MCAR test, to determine whether values were missing (completely) at random. Then we compared the main baseline characteristics of participants with and without missing data, to determine if there were any relevant differences between the groups. We compared the characteristics both visually and statistically with independent sample t tests and Mann–Whitney U tests.
Clinical course
The clinical course of cervical radiculopathy at 6 and 12 months was described using descriptive statistics. We used complete-case analyses to determine the clinical course of cervical radiculopathy.
Prognostic models
Multiple imputation methods were performed on the predictor and outcome measures with missing values. We used the Multivariate Imputation by Chained Equations (MICE) method with linear method imputation, and the number of imputations was related to the amount of missing data [13, 14, 20]. Demographic variables, predictor variables and the 6- and 12-month outcome variables were included in the imputation models [20].
We performed multivariable logistic regression analyses for each primary and secondary outcome in the imputed dataset. A priori, we aimed to include six factors in our models. The common rule of thumb states that the sample size for multivariable regression should be approximately 10 events in the smallest group per factor included in the analyses [13]. Therefore, we aimed to include a minimum of 60 participants in the smallest group (i.e., either recovered or non-recovered at 12 months) [13]. However, the final dataset was smaller than anticipated, because of the strict criteria we used to diagnose cervical radiculopathy. The recruitment period could not be extended, but initiatives were taken to maximize enrollment of suitable patients within the predefined time frame. This restricted the number of possible predictors per outcome. Because it was difficult to determine the three most relevant predictors for each outcome based on theoretical plausibility, we decided to include all six predefined predictors and apply strict bootstrapping techniques to correct for overfitting.
We used a manual backward selection procedure in the pooled analysis model, in which the factor with the highest significance level was removed, until all variables in the model had a p value < 0.157 [13, 21]. The predictive influence of the predictor was estimated by the odds ratio (OR). Performance of the model was determined by the explained variance and the accuracy of the model. The explained variance is described in terms of the Nagelkerke R2. The accuracy of the prognostic models was determined by the area under the curve (AUC). An AUC < 0.6 means that the prognostic model has no discriminatory value, an AUC > 0.8 reflects good discriminatory value [22]. Since no universal method has been described, the pooled AUC and Nagelkerke R2 were acquired by determining the median of the individual AUCs and Nagelkerke R2 of the imputed datasets [14].
The internal validity of the models was assessed through bootstrapping techniques with 500 repetitions. Bootstrapping is the preferred method for internal validation to determine the optimism in the initially developed model, based on the model’s performance in numerous (i.e., 500) bootstrap samples derived from the complete dataset. It determines a shrinkage factor that can be used to adjust the regression coefficient and performance indicators to correct for any optimism and to better reflect the actual performance of the model [13]. The models were internally validated in terms of explained variance and accuracy. The statistical analyses were performed in IBM SPSS, version 23 (IBM Corp, Armonk, NY, USA) and the bootstrap techniques in R statistics. All methods are reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) guideline [13, 14].