Predicting chronic pain after major traumatic injury

2.1 Study population and study setting

We recruited a convenience sample of patients admitted to Harborview Medical Center after traumatic injury, who underwent surgery for trauma. Harborview Medical Center is a 450-bed level-1 adult trauma center covering the five State Pacific Northwest region (Washington, Wyoming, Alaska, Montana and Idaho), with over 5,000 annual trauma admissions. The study was approved by the University of Washington Institutional Review Board. All patients gave informed consent electronically by signing their names on the consent form using the REDCap program (Research Electronic Data Capture) [27] on a hand held tablet.

Patients admitted from July 1, 2016 to December 31, 2016 were considered for enrollment if they were admitted as a trauma, according to the inclusion criteria of the Washington State Trauma Registry (Supplementary material File 1). All patients were admitted directly through the emergency department. Patients were screened via chart review every 24 h based on a list compiled by the trauma registry and approached in person by the study staff while they were hospitalized.

Patients who were less than 18 years old, unable to provide informed consent, in judicial custody, did not undergo surgery or were discharged from the emergency room or post-anesthesia care unit without admission, were excluded. Study data were collected and managed using REDCap [27].

Participation in the study did not involve any novel treatment regimens. All study participants received care according to the standards of the University of Washington Harborview Medical Center.

2.2 In-hospital assessment

Patients admitted to the hospital for surgical management were interviewed in person by the study staff after undergoing surgical intervention. We recorded gender, age, body mass index and race from the inpatient medical records. We recorded the following injury characteristics: injury severity score (ISS, range 0–75); mechanism of injury as defined in the Washington State Trauma Registry (fall, motor vehicle collision, pedestrian accident, other transportation, firearms/machinery/objects, and other); and injury at body regions with an Abbreviated Injury Scale (AIS) of at least 2 (scale 1–6), indicating a moderate or more severe injury [28].

We recorded the following variables of pre-trauma condition: presence of pre-existing pain, defined as yes/no; in case of pre-existing pain, pre-injury opioid use, defined as yes/no; the NIDA (National Institute of Drug Abuse) Assist Measure, a validated tool to screen for illicit drug use in general medical settings [29]; the PROMIS (Patient-Reported Outcomes Measurement Information System) Measures Computer Assisted Test for Depression, Anxiety, Physical function, Pain interference and Pain Behavior, reported as a T-score that is standardized to the population [30]; the Widespread Pain Index and Symptom Severity Score [31], reported as individual scores. For pre-trauma variables, we asked patients to recall their condition prior to hospitalization.

For in-hospital clinical characteristics, we recorded the following variables: intensity of post-operative pain at the site of injury 6 h after surgery, defined as numeric rating scale 0–10 (0=no pain, 10=worst pain imaginable); last current pain score before discharge (0–10); PROMIS Measures for Depression, Anxiety, Physical function, Pain interference and Pain Behavior; the Widespread Pain Index and Symptom Severity Score during hospitalization.

Regarding in-hospital clinical management, we recorded the following variables: type of surgery; number of surgeries, with surgery defined as treatment provided in the operating room necessitating care of an anesthesiologist; number of surgeries per patient; length of surgery; in-hospital opioid consumption, defined as morphine equivalent dose including parenteral and oral opioids; number of patients that used intravenous patient-controlled-analgesia; post-operative non-opioid medications as documented in the medical records; and number of patients that used regional analgesia intra- and/or post-operatively, defined as a single injection or placement of a catheter. To calculate the morphine equivalent dose, we used the guidelines of the Centers for Disease Control and Prevention for the oral equivalent doses (https://www.cdc.gov/drugoverdose/resources/data.html) and the recommendation by Gammaitoni et al. [32] for the iv equivalent doses.

2.3 Clinical management during hospitalization

The analgesic regimen after admission was determined by the admitting service, in consultation with the acute pain service as needed. The common practice of the acute pain team is to optimize medication therapy using opioid and non-opioid analgesics as well as adjunctive non-medication therapy. Rehabilitation psychology providers and spiritual care team members provide services during weekday hours.

The analgesic regimen during surgery was at discretion of the anesthesiologist. Regional anesthetic techniques are uncommonly used as the primary anesthetic at our institution, given the emergent nature and possible length of surgery. Peripheral nerve blocks are most commonly placed after surgical intervention in the post-anesthesia care unit (PACU).

In the immediate post-operative period, pain is managed by the anesthesiology team. After discharge from the PACU, pain is managed by the admitting service, with consultation from the acute pain service as needed. Physical therapy and mental health services are provided as needed, at discretion of the admitting service and/or the acute pain service.

2.4 Follow-up assessment

A single follow-up assessment was initiated at 3 months after hospital discharge to determine the presence of chronic pain at the site of trauma. Discharge from the hospital was chosen as time zero of the follow-up. Given the variation in hospital length of stay and number of surgeries, we assumed that patients reached a consistent level of recovery at the time of discharge from hospital. Participants were first sent an automated email to the address we had on file, requesting completion of an online survey through REDCap^®. Non-responders were contacted via a personalized email providing a telephone number and asking for a good time to administer the survey. Subsequent follow-up occurred by a telephone call to a known telephone number. We also approached alternate contacts if provided for a maximum of three times. After three unsuccessful attempts, patients were categorized as lost to follow-up. Surveys received after 6 months were not considered.

We asked about chronic pain at the site of trauma, defined as pain of any intensity at the site of trauma within the past week. We recorded: pain intensity at the site of trauma (0=no pain, 10=worst pain imaginable); current treatment for pain including opioid medication, non-opioid medication, marijuana and non-pharmacologic modalities; scores in the PROMIS Depression, Anxiety, Physical Function, Pain interference, and Pain Behavior; as well as the Widespread Pain Index and Symptom Severity Score.

2.5 Statistical prediction

The statistical plan was determined a priori. Chronic pain at follow-up was defined as a binary variable, i.e. as the presence of pain at the site of injury as self-reported by the patient. We used follow-up surveys conducted between 3 and 6 months post-discharge because this was the timing of highest response. Potential predictors identified a priori were: age; sex; presence of pre-existing chronic pain; intensity of post-operative pain at 6 h; in-hospital opioid consumption; injury severity score (ISS); location of trauma, defined as body region; use of regional analgesia intra- and/or post-operatively; pre-trauma PROMIS Depression, Physical Function, and Anxiety scores; in-hospital Widespread Pain Index and Symptom Severity Score; number of post-operative non-opioid medications.

With regard to missing data, no potential predictors had missingness of >20%. However, missing data were missing not at random (MNAR) and were instead informative of post-surgical patient experiences. In other words, the absence of a value for a particular variable told us something about the patient. To account for the informativity of missing pain scores 6 h after surgery, pain score was transformed to an ordinal variable with a category for missing values. This approach is not appropriate for explanatory models but has been recommended for predictive models with informative missingness [33], [34], [35].

Our multivariable model predicting chronic pain 3–6 months post-discharge was constructed in using a stepwise selection approach. First, continuous candidate predictors were assessed using LOWESS (locally weighted scatterplot smoothing) plots without transformation and transformed accordingly. Associations between all candidate predictors and chronic pain were assessed using bi-variable logistic regression. Significance tests were based on the Mann-Whitney test, Fisher’s exact, or unpaired t-test with unequal variance as appropriate. A parsimonious final set of predictor variables was selected by discarding candidate variables with a p-value>0.2 on bivariate analysis. The remaining potential predictors were included in a multivariable model, and any variables that did not retain their significance at the p-value>0.2 threshold were dropped. A significance threshold of 0.2, rather than the more traditional threshold of 0.05 found in explanatory modeling, was selected to increase statistical power and avoid “underfitting” a model with an erroneously small number of variables [36]. After the final model was developed, a reduced model, based only on variables available in the electronic medical record, was run to understand the “value add” of variables taken from study-specific instruments. The variables available in the electronic medical record were pain score 6 h post-surgery, presence of a head injury with AIS of 2 or greater, use of regional analgesia, and number of non-opioid medications. A sensitivity analysis used a best-subsets approach and leaps-and-bounds algorithm appropriate for logistic regression to select a parsimonious final model from all candidate predictors.

For the assessment of model performance, we tested model calibration (how well predicted outcomes matched known outcomes) using a Hosmer-Lemeshow goodness-of-fit test with 10 groups and a calibration plot. To assess model discrimination (how well the model distinguishes between subjects with and without chronic pain), we calculated the apparent and optimism-corrected c-statistic (area under the curve) for both the final and final reduced model, and compared them using the Delong method [37]. The apparent c-statistic, or area under the curve, quantifies the accuracy of the prediction model in the study sample. The optimism-corrected c-statistic is an estimate of how well the prediction model will perform in a new sample. We also calculated the index of prediction accuracy (IPA). The IPA is a rescaled Brier score combining discrimination and calibration into a single value, with 1.00 representing a perfect prediction model and an IPA<0 indicating a useless model [38]. Operating characteristics (sensitivity and specificity) of the model were calculated for prediction cut-points at 0.25, 0.50, and 0.75. For internal validation (the reproducibility of results within the original study cohort), we repeated these assessments using a bootstrapped resampling with replacement method for 200 iterations and calculated the overfitting percentage. All statistical models were run in Stata 15 (Statacorp LP, College Station, USA).

3.1 Study flow

From July 1, 2016 to December 31, 2016, 208 patients were approached by the study staff and 183 (88%) consented. Of 183 recruited patients, 173 (96%) completed the baseline interview. One hundred fifteen (66%) completed the follow-up survey within 3–6 months, with a mean of 127.5 days and a standard deviation of 30.0 days. Of the 115 patients who completed the 3–6 months survey, three patients did not answer the question regarding chronic pain and were excluded from the prediction model.

3.2 Clinical characteristics and management

Demographic and clinical characteristics for participants with complete baseline data and for participants with completed follow-up at 3–6 months are presented in Table 1. Table 2 displays demographic and clinical characteristics for participants with complete baseline data and for all patients admitted to Harborview Medical Center during the same time period and who underwent surgery, as recorded in the Harborview Medical Center Trauma Registry. In-hospital management is presented in Table 3. In our final cohort for prediction, 66% of patients had pain at the site of trauma 3–6 months after hospital discharge (74 of the 112 patients who answer the question regarding chronic pain). The clinical characteristics at follow-up are presented in Table 4.

3.3 Prediction model

Table 5 shows the results of our initial bivariate and multivariable models for all candidate predictors. The final prediction model (Table 6) included post-surgical pain score at 6 h, presence of a head injury, use of regional analgesia, PROMIS depression score, Widespread Pain Index score, and the number of post-operative non-opioid medications used for pain relief.

The apparent area under the ROC curve (AUROC) of 0.78 for predicting pain 3–6 months post-discharge was optimism-corrected to an AUROC of 0.73 (Fig. 1). The Hosmer-Lemeshow test showed acceptable fit, as did the calibration plot (Fig. 2). The model overfitting was 7.3%. The IPA for the final model was 0.233.

Fig. 1:

Discrimination of prediction scores for post-discharge chronic pain at 3–6 months, using study variables significant at p<0.2 and a reduced set of significant variables drawn from the electronic health records (EHR). The apparent area under the ROC curve (AUROC) based on significant study variables was 0.78, and optimism-corrected to 0.73. The AUROC based on a reduced set of EHR variables was 0.76, and optimism-corrected to 0.72.

Fig. 2:

Calibration of prediction scores for post-discharge chronic pain at 3–6 months.

The reduced final model, using only data available from the electronic health records, included post-surgical pain score at 6 h, presence of a head injury, use of regional analgesia, and the number of post-operative non-opioid medications used for pain relief. The apparent AUROC of 0.76 was optimism-corrected to 0.72 (Fig. 1). The difference between the final and reduced final models was not statistically significant (p-value=0.44). The IPA for the reduced final model was 0.157, 7.6 percentage points lower than the model that included measures of depression and WPI.

The sensitivity and specificity of the model at prediction cut-points of 0.25, 0.5, and 0.75 is presented in Table 7. The sensitivity analysis using the best-subsets approach chose use of regional analgesia, PROMIS depression score, and the use of ketorolac as the best multivariable model, based on penalized Akaike information criterion (AIC). This model had an apparent AUROC of 0.70 (Fig. 3) and an IPA of 0.134.

Fig. 3:

Discrimination of prediction scores for post-discharge chronic pain at 3–6 months, using study variables significant at p<0.2 and study variables chosen by vselect procedure and penalized Akaike Information Criterion (AIC). The apparent area under the ROC curve (AUROC) based on significant study variables was 0.78, and the AUROC based on vselect was 0.70.

4.1 Pain and co-morbidities at follow-up

We found that 66% of patients in our sample of trauma patients undergoing surgery had pain at the site of trauma 3–6 months after hospital discharge. Opioids were used in 28% of patients after 3–6 months post-discharge. This represents an increase from the 16% opioid use before the injury (Tables 1 and 4). Thirty percent of patients used marijuana for their pain after 3–6 months. There was a worsening in pain behavior, pain interference and physical function from baseline to follow-up (Tables 1 and 4).

Changes in pain-related management and co-morbidities were not the main aim of the present study. We therefore did not test for statistical significance. However, research in this area has been sparse and our findings have public health relevance. The data indicate that traumatic injury can lead to an increase in opioid use, so that previously opioid-naïve patients may become chronic opioid users. This finding is concerning in the light of the doubts on the long-term efficacy of opioids and concerns about their safety [22], [23], [24], [25], [26]. Because of the very limited evidence of efficacy of marijuana in chronic pain, also the finding of 30% use to treat pain at 3–6 months is concerning and deserves further investigation. As we did not record pre-injury marijuana use, we cannot however determine the number of new marijuana users in our cohort. It is unclear whether the decrease in physical function at follow-up is the result of pain, trauma, surgery or a combination of these factors. Investigating physical recovery after trauma and its relationship with pain is another relevant avenue of future research.

4.2 Prediction model

Our final predictive model included post-surgical pain score at 6 h, presence of a head injury, use of regional analgesia, PROMIS depression score, Widespread Pain Index score and the number of post-operative non-opioid medications used for pain relief. Our prediction model performed well, with an apparent area under the ROC curve (AUROC) of 0.78 that was optimism-corrected to 0.73. Although the index of prediction accuracy was better for the full model than the reduced, there was no statistically significant difference in the predictive ability between the full model and the model that included only variables that are routinely collected as part of the medical records. Taken as a whole, this suggests that patients at risk of chronic pain can be identified (1) with reasonable accuracy and (2) without the use of variables that are typically employed for research purposes.

While our predictive approach precludes inference on any single predictor [34], the majority of variables in our model predicting pain have been identified also in previous studies as important factors in the development of chronic pain. The presence of pain score at 6 h post-operatively as an element of the model is consistent with prior literature on chronic post-surgical pain, whereby high levels of acute pain were associated with chronic pain [13], [39], [40]. The area of injury (head) contributed to our predictive model. This finding cannot be interpreted in the frame of a predictive study and requires confirmation by future studies. The presence of regional analgesia among the predictors of chronic pain may be the result of the clinical context and the criteria for offering the treatment to patient. Patients with most severe acute pain were more likely to receive regional analgesia and may have been also more likely to develop chronic pain. The inclusion of depression in the predictive model reflects the well-established notion that patients with depressive symptoms are more prone to adverse pain outcomes. The predictive value of the widespread pain index is consistent with prior literature on the association between widespread pain index and poor pain outcomes [41], [42]. Finally, the finding on the number of non-opioid pain medications is again difficult to explain in a predictive model and requires investigations that specifically address the contribution of acute pain management to the development of chronic pain.

4.3 Strengths and limitations

To our knowledge, this is the first study on chronic post-injury pain that considered a comprehensive set of potential predictors, including pre-injury, in-hospital and clinical management variables. The prospective study design allowed for inclusion of variables not routinely collected on trauma patients and prevented recall bias for most variables. The retention at follow-up was challenging but could achieve the acceptable rate of 66% of completed follow-up data at 3–6 months, of the number of participants with complete baseline data. Follow-up variables were not limited to pain but included variables of pain-relevant comorbidities.

A larger patient population would have allowed the analysis of a larger set of potential predictors. As with any study of a convenience sample of patients, selection bias may limit the generalizability of findings to the broader trauma population. Based on available data (Table 2) we believe that our results are generalizable, pending external validation. In this regard, our model needs to be validated in other populations to confirm generalizability. However, our focus on variables that are routinely collected in electronic medical records implies that a multi-site validation should be feasible, and our model can be easily implemented by any provider with access to medical records.

The predictive rather than explanatory nature of our model prevents the interpretation of the contribution of individual factors. Part of the pre-injury variables had to be collected during hospitalization, relying on patients’ recall of their pre-injury state. While this was unavoidable, the validity of this approach is uncertain.