Identifying individuals with chronic pain after knee replacement: a population-cohort, cluster-analysis of Oxford knee scores in 128,145 patients from the English National Health Service

Data were obtained from the patient-level, anonymized English NHS Patient-Reported Outcome Measures (PROMs) programme available from the NHS Digital website [9]. All patients undergoing KR through the English NHS, or on behalf of it, are eligible to be included in the programme. Two questionnaires are sent to participants before and after surgery by the PROMs programme, a disease-specific and a generic quality of life measure: the OKS [10] and the EQ-5D [11]. For this analysis, data were extracted for UK financial years 2012 through 2015.

Patients undergoing primary KR, with complete post-operative OKS were included in the present study. There were no restrictions on age or sex. The dataset did not include the indication for or type of surgery performed, hence records were included regardless of the underlying condition or type of KR. As the unit of analysis was KRs and records were fully anonymised, it was not possible to determine if patients had their contralateral knee replaced within the same period of analysis. It was therefore possible for patients to be included twice in the dataset.

The OKS is a 12-item score which measures knee pain, stiffness, and functional disability within the previous 4 weeks [10]. Responses are scored using a 0–4 Likert scale. Item scores are summed to calculate a total score ranging from 0 (worst possible score) to 48 (best possible score). The PROMS programme collects post-operative data 6 months after surgery. Chronic post-surgical pain is widely accepted to be pain of at least three to 6 months duration that develops or increases in intensity after a surgical procedure and significantly affects health-related quality of life [12, 13]. This is applicable to KR, as most pain relief is obtained in the first three to 6 months [14, 15] and continuing pain at 6 months after surgery is a cause of dissatisfaction [16].

Previous work has identified pain- and function-related subscales within the OKS [17], with further details given in Additional file 1. A raw OKS pain subscale (OKS-PS) summary score can be calculated by summing the responses of its seven items, hence ranging from 0 (most pain) to 28 (least pain). Although the OKS, and particularly the pain subscale, could be considered an unidimensional measure of highly-correlated items, each item captures a different dimension of pain. Some pain items are highly intertwined with function, reflected by the inclusion of two items in both the pain and function components [17]. Our analysis allowed for each item to independently contribute to the generation of pain severity clusters.

To identify the high-pain group, hierarchical clustering was used to group patients into clusters based on the similarity of their OKS-PS item scores. Hierarchical clustering is commonly used, easily interpretable and has been applied in several previous studies on knee pain in order to identify pain profiles, and broader subgroups which are then examined for correlation with pain [18‐21]. Cluster analysis splits a set of participants into groups or clusters, so that participants within the same cluster are most similar, and any two clusters are as distinct as possible from one another. Agglomerative hierarchical clustering follows a sequence whereby clusters are repetitively merged until a pre-specified number of clusters is achieved, or all participants have been merged into one single cluster. We conducted agglomerative clustering with all patients beginning as individual clusters, which were then successively merged into all possible numbers of clusters, starting at two (the natural minimum) up to 28 clusters (the maximum, where each cluster would correspond to each possible value of the OKS-PS). This constituted the essence of our data-driven approach, considering all possibilities allowed by the data instead of setting discretionary limits on the number of clusters to be identified. We assumed that it is not known, a-priori, if subgroups of KR patients exist based on their levels of pain and, if so, how many there are. By applying a data-driven approach, we allowed the model to answer these questions based on the features of our population.

For each set of clusters (2 to 28), the distribution of the OKS-PS corresponding to the highest-pain cluster, i.e. that reporting the lowest values, was examined. This was used to derive a cut-off point, defined as the highest value taken by the highest-pain cluster on the OKS-PS. Changes in the cut-off point as the number of clusters varied are reported. Further details on the methodology have been published previously [22].

Based on the cut-off point identified in the primary analysis, patients were placed into either a ‘high-pain’ or a ‘low-pain’ group, and scores reported for the overall OKS, OKS subscales, OKS-PS items, EQ-5D Visual Analogue Scale (VAS), and the EQ-5D summary score comprising mobility, self-care, usual activities, pain/discomfort, and anxiety/depression. EQ-5D dimensions are scored using a 1–3 Likert scale [11] and a summary index calculated by applying UK general population preference weights [23]. The summary score is anchored at 0 signifying ‘death’ and 1 ‘perfect health’, negative values being possible and interpreted as health states worse than death. Answers to two questions about the results and impact of the surgery are also reported by group. Analyses were performed using Matlab R2015 and Stata.

Figure 1 shows the distribution of different-size clusters over the OKS-PS summary score from the primary analysis. When the clustering algorithm derives k = 2 clusters, the OKS-PS distributions for the two clusters have a large overlap. As the number of derived clusters increases, the degree of overlap in their corresponding OKS-PS distributions reduces and both the mode and the upper limit shift. Figure 2 illustrates this shift in the resulting cut-off point, in particular how it decreases as the number of clusters increases from k = 2 to k = 4. However, for k ≥ 4, the cut-off point remains constant at OKS-PS = 14, implying that the OKS-PS score range of 0–14 characterises the high-pain cluster whether the sample is split into four or more clusters. Using OKS-PS ≤ 14 to identify participants in chronic pain led to a sensitivity of 100% and specificity of 89%.

The cut-off points obtained through secondary analysis were different to that obtained through the primary analysis. As Fig. 2 shows, running 100 random re-orderings and applying hierarchical clustering led to a mean cut-off point at k = 4 clusters of OKS-PS = 18.5 (SD = 2.6), whilst when k-means was applied the mean cut-off was OKS-PS = 14.7 (SD = 0.9). The Figure also illustrates that these cut-off points are not constant with increasing numbers of clusters. At OKS-PS = 18.5, sensitivity and specificity were 100% and 74%, whereas at OKS-PS = 14.7 they were 100% and 89%, respectively.

Using the OKS-PS cut-off of 14, a total of 18,522 out of 126,064 (14.7%) patients with KR were found to be in the high-pain group. Table 2 shows the demographics and outcome scores for the high- and low-pain groups.

Age and sex were largely similar between both groups, with the exception of a larger proportion of 50–59 year olds and lower proportion of 70–79 year olds in the high-pain compared to the lower-pain group. Pre-operatively, patients in the high-pain group reported worse pain, function and health-related quality of life. Post-operatively, outcomes were also poorer for the high-pain group. A higher percentage of the high-pain group judged the outcome of their operation as ‘fair’ or ‘poor’, and perceived little or no improvement in their symptoms from their pre-operative status. Figure 3 shows how the high-pain group consistently reported greater pain, in average, in all OKS-PS items. The greatest differences were reported in pain severity, limping and night pain. The overall improvement of the standardised OKS-PS (scale of 0–100) was 46 for the lower-pain group and 10 for the high-pain group. Outcomes in function were even worse for the high-pain group, with a mean improvement of 2 points compared to 27 for the lower-pain group. Post-operative health-related quality of life was also worse in the high-pain group, with lower scores by approximately the same magnitude across all five dimensions. As shown in Fig. 4, however, the median of the EQ-5D summary score for the high-pain group saw a larger improvement after surgery than the lower-pain group, but the middle half of the high-pain group remained essentially within the same range of scores reported pre-operatively. Notably, 17% of the high-pain group reported post-operative EQ-5D summary scores lower than zero, compared to less than 1% for the lower-pain group.