Anterior surgical treatment for cervical degenerative radiculopathy: a prediction model for non-success

This is a multicentre longitudinal study following the recommendations for reporting in observational studies, STROBE criteria [44] and the methodological framework proposed by the PROGRESS framework [33, 40]. The manuscript is reported according to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines [9]. We used the Prediction model Risk Of Bias Assessment Tool (PROBAST) to minimize the risk of bias [27, 45]

Our research protocol was approved by the Norwegian Committee for Medical and Health Research Ethics Midt (2014/344). Written informed consent was obtained from all patients.

Data from the Norwegian Registry for Spine Surgery (NORspine) from 2011 to 2016 was used. NORspine is a government-funded comprehensive clinical registry receiving no industry funding and used for quality assessment and research. Informed consent is obtained from all patients before they enter the registry. Currently, all centres performing cervical spine surgery in Norway report data to NORspine (coverage = 100%), and the operation recording rate is 78% (completeness) [3]. Patients who had undergone anterior cervical discectomy and fusion or arthroplasty surgery due to cervical degenerative radiculopathy in the period were included. For both groups, baseline characteristics and 12-month outcome data were similar, except from baseline Neck Disability Index and neck pain scores, which were slightly higher in the arthroplasty group (p = 0.02 and p = 0.002, respectively). Also, arthroplasty patients were operated in significantly lower number of levels (p < 0.001). Patients undergoing posterior cervical procedures due to CDR, as well as all patients operated for myelopathy symptoms, were excluded. Patients operated for tumours, fractures and primary infections are not included in NORspine.

Patients completed data at admission for surgery (baseline), after 3 and 12 months. Surgeon’s forms containing information about diagnosis, treatment and comorbidity were completed during the hospital stay. Only cohort participants responding to the 12-month questionnaire were included in the present study. Follow-up was conducted by the central registry unit without involvement of the treating hospital. The patients responded by questionnaires sent and returned by mail. One reminder with a new form was sent to non-respondents within 2 weeks.

For the primary outcome, the Neck Disability Index (NDI) (0–100), non-success was defined as an absolute score of > 26 at 12-month follow-up. For the secondary outcome, arm pain intensity assessed by numerical rating scale (NRS-AP) (0–10), non-success was defined as a score ≥ 3 at 12-month follow-up. These estimates are based on a previous study for patients undergoing CDR surgery in Norway [25].

Candidate predictors for non-success were selected from the comprehensive NORspine questionnaire administered before surgery, which consists of information about sociodemographic factors, lifestyle, work and clinical variables in addition to patient-reported outcome measures (PROMs). Data from the surgeons’ forms were used for information about diagnosis, treatment, comorbidity, the American Society of Anaesthesiologists physical status (ASA), surgical indication and type of operation. The selection of the final set of predictors was made after a thorough literature review where we identified the factors that have been found to be significantly and consistently associated with outcomes after CDR surgery.

The following predictors were selected for the model: gender, age groups (below 40 years, between 40 and 60 years or above 60 years), work status prior to operation (on sick leave, retired or disabled; on rehabilitation, out of work or on work return training; student, fully working or housewife/househusband), physical demands in work (working with computers, sitting, light physical work or hard physical work), educational level (high school or less, less than 4 years university or 4 or more years of university), mother tongue (Norwegian or non-native speaker), pending litigation (yes/no) (pending litigation defined as unresolved claims or litigation issues against the Norwegian Public Welfare Agency Fund concerning permanent disability pension or compensation claims against private insurance companies or the public Norwegian System of Compensation to Patients), duration of arm pain (less than 3 months, 3 to 12 months or more than 12 months), duration of pre-operative paresis (no paresis, less than 3 months or more than 3 months), body mass index (BMI) (equal to or below 30 or above 30), smoking (yes/no), comorbidities (yes/no), previous neck surgery (yes/no), number of surgical levels (one or more than one), daily use of analgetic drugs (yes/no), ASA level (level 1–2 or level 3–4), arm pain neck pain ratio (above 1 or below or equal to 1) [30] and anxiety/depression by the item on the EuroQol-5D-3L (EQ-5D) questionnaire (“moderate” to “extremely” anxious or depressed or “not anxious or depressed”). In addition, baseline outcome scores were included as potential predictors; the baseline scores were categorized into low, medium and high by percentile distribution.

Since no consensus on sample size in prognostic modelling exists, we chose to follow the recommendations by Steyerberg [38]: (a) aiming for at least 100 events as a minimum for reliable estimation of the average risk and (b) aiming for at least 10 events per variable (EPV) and preferably 20, for reliable prediction modelling if the event rate is < 20% and higher EPV values if the event rate is between 20 and 80%. In the present material, approximately 700 cases had non-success at 12-month follow-up, and with this large number of EPV, we had nearly 40 cases per event and good statistical power for the prediction model analyses. The large EPV will reduce the potential for overfitting and optimism of the final models. Overfitting is defined as fitting a statistical model with too many effective degrees of freedom in the modelling process. Estimation bias is defined as the overestimation of effects of predictors because selection of the effects withstood a statistical test, whereas optimism is defined as the difference between true performance (performance in the underlying population, e.g. external validation sample) and apparent performance (development sample) [38].

Statistical analyses were performed using IBM SPSS Statistics version 26 for Windows and the STATA version 16 for Windows. Missing data was checked for all variables and are reported together with descriptive data. Frequencies were used for categorical data and mean and standard deviation (SD) for continuous data. Continuous variables, such as baseline disability, were categorized to be adapted into a risk matrix. The distribution of baseline and 12-month scores of the two outcome measures are presented by mean scores and SD.

First, a univariable analysis of the candidate predictors was performed to assess the crude association between each candidate predictors and the two outcomes. Associations between outcomes and predictors are expressed as odd ratios (OR) with a 95% confidence interval (CI). Predictors reaching p < 0.1 in these analyses were entered into two multivariable logistic regression models (for primary and secondary outcome), where a stepwise backward elimination method was used. Variables that were not statistically significant (p > 0.05) in the multivariable models were removed from the model. The performance of the two final models was evaluated with (1) the explained variance by Nagelkerke’s R², (2) the Hosmer–Lemeshow test p > 0.05) and (3) the discriminative ability of the model (the likelihood that the model allocates higher predicted risks to patients who achieve non-substantial improvement and lower predicted risks to those who do not) assessed by calculating the area under the receiver operating curves (AUC), also often referred to as the c-index [41]. The larger the AUC, the greater is the discriminative ability of the model. The discriminative performance of the models was considered acceptable if the AUC was ≥ 0.7 and good if the AUC was ≥ 0.8 (the c-criterion).

Internal validation was conducted by a bootstrap procedure (1000 samples) to estimate the amount of optimism in the two final models [26, 39]. A slope value was calculated (the closer to 1.0, the less over-optimism) and used to correct and shrink the regression coefficients, the R² and the c-index.

We assessed the potential clinical utility of the final prediction model for non-success in neck disability by developing a risk matrix for two hypothetical patient case profiles with few and many predictors present. Regression coefficients from the final disability model were converted into probabilities, and a risk score for each of the two individual case profiles was calculated by the sum of the products of individual values of each predictor variable and its regression coefficient. Depending on the presence or absence of the risk factors, the matrix was then calculated as probability for a non-substantial improvement after 12 months for each of the patients.

In the current healthcare environment, value-based thinking has brought more focus on quality and appropriateness of care. Also, as degenerative neck surgery is becoming increasingly safe and efficient, there is a need for more knowledge about which patients are not improving from surgery. The two present models can be used in a clinical setting to predict which patients will benefit from a surgical intervention and who will be better off being treated conservatively. To exemplify how these models can be used in a surgical practice, we produced a risk matrix constituted of two hypothetical patient scenarios for disability; one where the patient had several of the risk factors and another where the patient had only a few risk factors (Table 4). The patients with few predictors had low probability for non-success (0.13), while several predictors involved a high risk for non-success (0.92). According to our model, a patient with similar characteristics and symptomatology as described in the case study with few predictors should be reassured that surgery is a safe option in terms of improving from baseline arm pain and disability. Patients with a similar clinical picture as patient 2 with several positive predictors, on the other hand, should be counselled about alternative treatment strategies.

The present models can be further developed into a risk calculator to assess the probability of success or failure to achieve substantial change for every patient in a surgical practice. However, the model will first need to be further validated in other study populations. The feasibility of a risk calculator should also be evaluated.

An advantage of the present study is the large sample size of data captured in a national registry. NORspine was designed to prospectively capture important candidate predictors and PROMs prior to and during the year following surgery. The registry covers all the hospitals and private clinics conducting surgery on spinal disorders in Norway. A total of 78% of the operations are recorded in the registry [36]. Furthermore, our two models were well balanced with respect to the risk of overfitting, in particular the disability model which showed high accuracy with only seven included predictors.

In our study, we chose to include patients operated with both arthroplasty and fusion. The baseline characteristics and 12-month outcome data were similar between the groups, except for slightly higher NDI and NRS-NP scores for the arthroplasty patients at baseline, as well as a lower number of operated levels. There is no current consensus about the use of arthroplasty vs fusion in patients with degenerative cervical disease [8, 13, 16, 17, 42, 46]. One may question whether the results of the fusion group in our study can be generalized to the arthroplasty group since there are only 1% of arthroplasty patients in our cohort. Further studies are warranted to elucidate this issue.

Loss to follow-up was 35.6% at 12-month follow-up and could represent a selection bias. However, two recent Scandinavian spine registry studies based on similar cohorts have found that a loss to follow-up did not bias conclusions about treatment effects [22, 37].

Another potential limitation is related to the cut-off estimates of the applied PROMs. In the present study, we decided to use estimates of non-success instead of the concept of MCID. The main reason is that MCID often show to be less than measurement errors or estimates for smallest detectable change [43], making it difficult for a patient and/or a clinician to judge the clinical meaningfulness of these estimates. By using stricter estimates reflecting a substantial rather than minimal change, we argue that these cut-offs are better suited for use in the development of prediction models for non-success (or success).

The major limitation of the present study is that we did not externally validate the final models. External validation is necessary before these models can be further developed into a risk calculator used in clinical settings. Risk calculators may help inform discussions of surgical treatment options between surgeons and patients and lead to more accurate judgement of operative risk. In a clinical decision-making process, the probability of successful or non-successful outcomes of conservative treatment strategies also needs to be taken into consideration. The present study only investigated outcomes after surgical treatment and cannot be generalized to outcomes after non-surgical treatment options. Thus, there is a large need for exploring prediction models for both surgical and non-surgical treatment trajectories and outcomes.