Predicting severe pain after major surgery: a secondary analysis of the Peri-operative Quality Improvement Programme (PQIP) dataset

Acute postoperative pain is common, distressing and associated with increased morbidity. Targeted interventions can prevent its development. We aimed to develop and internally validate a predictive tool to pre-emptively identify patients at risk of severe pain following major surgery. We analysed data from the UK Peri-operative Quality Improvement Programme to develop and validate a logistic regression model to predict severe pain on the ﬁ rst postoperative day using pre-operative variables. Secondary analyses included the use of peri-operative variables. Data from 17,079 patients undergoing major surgery were included. Severe pain was reported by 3140 (18.4%) patients; this was more prevalent in females, patients with cancer or insulin-dependent diabetes, current smokers and in those taking baseline opioids. Our ﬁ nal model included 25 pre-operative predictors with an optimism-corrected c-statistic of 0.66 and good calibration (mean absolute error 0.005, p = 0.35). Decision-curve analysis suggested an optimal cut-off value of 20 – 30% predicted risk to identify high-risk individuals. Potentially modi ﬁ able risk factors included smoking status and patient-reported measures of psychological well-being. Non-modi ﬁ able factors included demographic and surgical factors. Discrimination was improved by the addition of intra-operative variables (likelihood ratio v 2 496.5, p < 0.001) but not by the addition of baseline opioid data. On internal validation, our pre-operative prediction model was well calibrated but discrimination was moderate. Performance was improved with the inclusion of peri-operative covariates suggesting pre-operative variables alone are not suf ﬁ cient to adequately predict postoperative pain.


Introduction
Acute postoperative pain is common, with up to 47.2% of patients reporting severe pain within the first 24 h of surgery [1,2]. It is associated with a negative patient experience and may also be associated with respiratory and/or cardiac complications, prolonged hospital stay, limited or delayed return to normal activity and the development of chronic postsurgical pain [2][3][4].
Identification of the patient at increased risk of problematic postoperative pain is challenging, with potential risk factors spanning patient, anaesthetic and surgical domains. Patient factors may include nonmodifiable variables such as age and sex, as well as potentially modifiable characteristics such as anxiety, psychological distress and high levels of catastrophisation [5,6]. Additionally differences in pain perception, preoperative pain, pre-operative opioid usage and the presence of chronic pain have all been associated with problematic acute postoperative pain [5][6][7]. Anaesthetic and surgical factors may include the choice of analgesic regimen, inclusion of regional blocks, type and duration of surgery and the choice of surgical incision [5,6].
We hypothesised that postoperative morbidity that is related to acute, or acute-on-chronic pain, could be mitigated by better identification and pre-emptive holistic management of`at-risk´patients. The first part of this work is to attempt to identify patients at risk, ideally pre-operatively.
Identification of patients at high risk of developing postoperative severe pain pre-operatively allows for better pre-operative decision-making, for counselling and for the introduction of evidence-based interventions. From a study perspective, it also has the advantage of preceding any anaesthetic or surgical intervention which might introduce unmeasured confounding. However, to determine the relative importance of predictors of severe pain, the effects of healthcare processes and peri-operative events should also be taken into account.
We undertook a secondary analysis of the Perioperative Quality Improvement Programme (PQIP, www. pqip.org.uk) dataset examining potential risk factors for postoperative pain. We focused principally on preoperative variables with additional consideration of intraoperative factors to ascertain their relative contributions to acute postoperative pain. Our aim was to develop and internally validate a model to predict risk for patients undergoing major surgery.

Methods
The Peri-operative Quality Improvement Programme is a prospective, multicentre, observational cohort study established in 2016 which collects data on adult patients (aged ≥ 18 y on date of surgery) undergoing major, planned non-cardiac surgery in UK National Health Service (NHS) hospitals [ Figure S1.
Categorical variables were regrouped (levels collapsed) where classes contained few individuals or events, as reported in the Results section.
As this was a secondary data analysis, sample size was determined by the eligible PQIP cohort. Adequacy of sample sizes and numbers of events were assessed according to recommendations for the development of prediction models for binary outcome data [15] (see online Supporting Information Appendix S1 for calculation). then subtracted from the apparent accuracy of the initial model to get the optimism-corrected estimate [14].
Where the performance of models was compared, calibration was assessed as above. If both showed good calibration, discrimination (by c-statistic) was compared as were Brier scores (squared differences between actual binary outcomes and predictions). For nested models (i.e. where one model includes a subset of the predictors in another) the likelihood ratio test was used. This compares the goodness-of-fit of the models with the data and provides a chi-squared statistic (v 2 ) and p value based on the null hypothesis that the models are the same fit to the data.
Decision-curve analysis was used to describe and compare the clinical implications of using each model at different thresholds. In decision-curve analysis, a model is considered to have clinical value if it has the highest net benefit across the whole range of thresholds for which a patient would be labelled as`high risk´ [17]. We also calculated the optimal threshold according to the Youden index, a means of summarising the receiver operating characteristic curve by giving equal weight to sensitivity and specificity [18]. tidyverse. The analysis code is available online [19]. Our findings are reported in accordance with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement [20].

Results
The   Decision curve analysis revealed that a threshold of 50% predicted probability of severe pain is similar to treating nobody, and a threshold of 10% or lower is equivalent to treating everybody. The potential for net benefit with the model appears maximal with a threshold between 20% and 30% (Fig. 2b). The optimal cut-point according to the Youden index was slightly lower at 18%.
As described above, a subset of patients from 2019 onwards (n = 6388) had data on pre-operative opioid usage.  Information Table S7).    The patients included in the primary analysis all had additional peri-operative (intra-operative and recovery) and postoperative (24 h) data (Table 1) Table S8).
The full model specification (online Supporting Information  Figure S3a  Decision-curve analysis showed increased net benefit across the range of probabilities (online Supporting Information Figure S3b), but this does not account for the difference in time-points at which decisions would be made.

Discussion
In this secondary analysis of data from the PQIP database, we have developed and internally validated a prediction model for severe pain on postoperative day 1 after major, non-cardiac surgery which utilises only pre-operative patient data. Model performance was limited, but several potential contributory factors were identified. This is the first attempt to systematically develop a peri-operative pain prediction model using such a large, high-quality dataset in a mixed surgical population.
Severe pain occurred in 18.4% of patients. Whilst in keeping with the ranges reported in previous case series [1,3] this represents a significant minority of patients, with potential implications for patient outcomes given the associations between postoperative pain and morbidity [2].
We focused on pre-operative variables to increase the clinical utility of our model [10] with the ultimate aim that prediction would enable clinicians to take preventative rather than reactive approaches. Many of the key predictors were non-modifiable patient and surgical aspects, for example female sex and younger age, both of which have been reported in previous studies [11,12,21]; a past medical history of diabetes; and thoracic surgery. A potentially modifiable lifestyle factor identified was current smoking status, which has previously been cited [12]. We did not replicate the previously reported finding of baseline opioid usage as a risk factor [11,12].
We found baseline patient-reported outcome measures to be relatively strong predictors, particularly those around psychological symptoms of anxiety/ depression but also reported pain/discomfort. Anxiety and depressive symptoms, as well as higher levels of preoperative pain, have previously been recognised as contributors to the risk of problematic postoperative pain [11,12,21]. Alongside these are complex factors we were not able to assess, such as pain catastrophising, feelings of helplessness and higher than expected pain [11,21]. The importance of functional domains has also previously been highlighted [22]. Whilst our model did include a WHO Disability Assessment Schedule 2.0 measure of ability to undertake normal activities, the EQ5D domain of`usual  activities´was not included. These patient-reported symptoms represent potential areas of modifiable benefit if a holistic approach is taken to peri-operative assessment and pre-optimisation before major surgery.
The factors most strongly associated with reduced risk of pain were all surgical factors: speciality, mode and grade of surgery. Whilst these speciality-specific factors might not be directly modifiable by the anaesthetist, they could contribute to a shared decision-making approach with both surgeons and patients, in which their potential impact on postoperative course could be explored [10].
Whilst predictions based purely on pre-operative data might have optimal utility in clinical practice [10], the discriminatory performance of the model was limited, though a c-statistic of 0.66 is similar to other reported predictive models [3,11]. One of the largest existing analyses is a model developed using data from approximately 50,000 patients in the international PAIN OUT registry [11]. However, this model included patients from 2011 to 2015, and the express aim was to produce a simple scoring tool, patient reported outcome measures were only recorded postoperatively rather than at baseline.
Additionally, the most important predictor for severe postoperative pain found was the country in which the surgery took place [11], our analysis has the advantage of including patients from only one healthcare system.
Taken together, our primary model and existing analyses suggest that pre-operative data alone are not sufficient to accurately predict which patients will go on to experience problematic acute postoperative pain. We   We have shown that data collected as part of a prospective cohort study can be used to develop a tool for predicting severe pain on postoperative day 1 with limited performance, but which serves to highlight several potentially important and modifiable contributing factors.
However, it is likely that several important patient and process measures are missing from the available dataset and that pre-operative factors alone are not sufficient for accurate prediction beyond risk-stratification. Further work will be required to explore additional factors which might improve predictive performance and assess how the tool might be applied in clinical practice.

Supporting Information
Additional supporting information may be found online via the journal website.
Appendix S2. Members of Peri-operative Quality Improvement Programme (PQIP) delivery team and collaborative. Figure S1. Non-linear relationships between predictors and outcome.