Predicting long-term postsurgical pain by examining the evolution of acute pain

Abstract

Background:

Increased acute postoperative pain intensity has been associated with the development of persistent postsurgical pain (PPP) in mechanistic and clinical investigations, but it remains unclear which aspects of acute pain explain this linkage.

Methods:

We analyzed clinical postoperative pain intensity assessments using symbolic aggregate approximations (SAX), a graphical way of representing changes between pain states from one patient evaluation to the next, to visualize and understand how pain intensity changes across sequential assessments are associated with the intensity of postoperative pain at 1 (M1) and 6 (M6) months after surgery. SAX-based acute pain transition patterns were compared using cosine similarity, which indicates the degree to which patterns mirror each other.

Results:

This single-center prospective cohort study included 364 subjects. Patterns of acute postoperative pain sequential transitions differed between the “None” and “Severe” outcomes at M1 (cosine similarity 0.44) and M6 (cosine similarity 0.49). Stratifications of M6 outcomes by preoperative pain intensity, sex, age group, surgery type, and catastrophizing showed significant heterogeneity of pain transition patterns within and across strata. Severe-to-severe acute pain transitions were common, but not exclusive, in patients with moderate or severe pain intensity at M6.

Conclusions:

Clinically, these results suggest that individual pain-state transitions, even within patient or procedural strata associated with PPP, may not alone offer good predictive information regarding PPP. Longitudinal observation in the immediate postoperative period and consideration of patient- and surgery-specific factors may help indicate which patients are at increased risk of PPP.

1 |. INTRODUCTION

Over 60% of the 100 million patients worldwide who have surgery each year will suffer from moderate to severe postoperative pain, and increased acute postoperative pain intensity has been associated with the development of persistent postsurgical pain (PPP) in mechanistic and clinical investigations.(Apfelbaum et al., 2003; Buvanendran et al., 2019; Buvanendran et al., 2015; Coderre et al., 1993; Gan et al., 2014; Gilron et al., 2017; Perkins and Kehlet 2000) Prior work has further refined these associations and suggests that the acute postoperative pain intensity trajectory is associated with the risk of developing PPP.(Gilron et al., 2017; Kampe et al., 2017; Katz et al., 1996; Kehlet et al., 2006) Importantly, the methodology in these studies has relied on aggregates or daily average pain intensities rather than information contained within repeated assessments of pain intensity within individual days.(Althaus et al., 2014; Althaus et al., 2012; Chapman et al., 2011; Chapman et al., 2012; Lavand’homme et al., 2014) While offering several advantages with respect to the standardization of data acquisition and statistical analyses, one tradeoff is that such data collection and analytical strategies may miss relevant signals resulting from the hour to hour changes inherent to postoperative recovery.

Despite intuition and evidence suggesting the existence of latent information within hour by hour changes, the analysis of pain intensity ratings at time intervals more granular than daily averages in the acute postoperative setting poses several statistical challenges given the irregular intervals of assessment.(Geurts 2001; Keogh et al., 2006a; Marsh and Shibano 1984) Moreover, translating the information contained within complex time series to clinically interpretable information poses further difficulties and limits the utility of otherwise widely available clinical data.(Aigner et al., 2007; Keogh 2006; Keogh et al., 2006b) To address these challenges, our team has previously examined these patterns as transition matrices in the form of Markov Chains.(Tighe et al., 2016) We further refined this line of investigation using symbolic aggregate approximation (SAX) as a method to display “transition icons” of acute postoperative pain data, while also providing more statistically robust comparisons among patient strata than available in Markov Chain transition matrices.(Kumar et al., 2005; Liao 2005; Lin et al., 2007; Shieh and Keogh 2008; Tighe et al., 2017)

SAX representations join a number of time series representation methods that aim to characterize time series data from different perspectives.(Lin et al., 2007; Malinowski et al., 2013) SAX offers several advantages for clinical pain intensity time series, including greater precision in non-periodic data than spectral decomposition methods, low computational complexity, and lower bounding.(Ding et al., 2008) Lower bounding is especially important because it indicates that the SAX representation of the time series accurately reflects the original data, even in very large collections of time series data.(Faloutsos et al., 1994) Clinically, SAX representations may be appealing given the visual intuition of the SAX-based transition icons that can summarize and highlight important findings within the data in a single, easy to interpret graphical manner.(Duan et al., 2016; Georgoulas et al., 2015; Karvelis et al., 2015; Keogh et al., 2006b)

While our prior work highlighted differences in SAX representations of acute pain intensity time series among differing demographic and procedural groups, the lack of longer-term outcomes limited its theoretical value for extrapolation to individual patients. To address this gap, we now examine how SAX representations differ according to PPP outcomes. Our primary aim was to examine differences in SAX representations based on clinically collected postoperative pain intensities from postoperative day (POD) 0 to 7 according to mean daily pain intensity at 1, and 6 months after surgery. The second aim was to examine for SAX pattern differences in age, sex, type of surgery, and catastrophizing at 6 months using acute pain intensity ratings from POD 0 to 7.(Fillingim et al., 2009; Gerbershagen et al., 2013; Gerbershagen et al., 2014; Khan et al.,2011; Riley III et al., 2000; Sullivan et al., 1995)

2 |. Methods

This prospective cohort study was approved by University of Florida Institutional Review Board-01 (protocol IRB201500153). All participants provided written informed consent prior to enrollment. These analyses were conducted in accordance with STROBE criteria.(Malta et al., 2010; Vandenbroucke et al., 2007)

2.1 |. Study participants

Study subjects were prospectively enrolled surgical patients scheduled for elective major orthopaedic, thoracic, colorectal, urologic, pancreatic/biliary, or spine procedures with an anticipated postoperative admission of at least 48 hours and life expectancy of at least 6 months. All subjects received surgery at UF Health Shands Hospital in Gainesville, Florida, USA, between November 2015 and May 2019. Patients were screened, recruited, and enrolled in the study from the presurgical clinic and all baseline sociodemographic variables were collected within 30 days prior to surgery. Patients undergoing colorectal, thoracic, pancreatic/biliary, and orthopaedic surgeries (but, notably, not spine surgeries) were generally offered neuraxial, paravertebral, or perineural regional anesthetics as adjuncts to general or neuraxial anesthesia as part of a multimodal analgesic strategy. These applications of regional anesthetics were conducted in line with established clinical care pathways routinely used for various surgical procedures at the study site, such that the use of these regional anesthetics were associated with the surgical procedure in question.

2.2 |. Data collection

Following enrollment, subjects were surveyed for preoperative catastrophizing using the Pain Catastrophizing Scale (PCS).(Sullivan et al., 1995) The (PCS) results were further binned at a cutoff score of 30 based on prior work suggesting clinical relevance of high catastrophizing at this level.(Sullivan et al., 1995) Preoperative pain intensity was assessed before surgery using the brief pain inventory (BPI).(Cleeland and Ryan 1994; Keller et al., 2004) Sociodemographics and clinical data including age, sex, and type of surgery were derived from the electronic health record system.

Following surgery, pain intensity data were collected by nursing staff according to clinical protocol using the Defense and Veterans Pain Rating Scale (DVPRS) with values ranging from 0 to 10. (Buckenmaier et al., 2013; Polomano et al., 2016) Pain scores were generally recorded approximately every 4 hours according to nursing protocol, with a repeat query within 1 hour following administration of analgesic medications. Patients in the intensive care unit or intermediate care unit were generally assessed every 1 to 2 hours depending on acuity level. When clinical staff documented the pain intensity of “patient asleep,” this value was converted to missing in this study to reflect the absence of reported pain intensity. All pain scores were recorded in the EPIC electronic health record (EHR) system with a corresponding date/time stamp; notably, the EPIC EHR system also provided guidance on how to score the DVPRS at the time of data entry, and includes validation of entries. All pain intensity rating times were indexed to the definition of end of surgery indicating that the surgical dressing had been applied and that surgical tissue trauma had theoretically ceased. Pain intensities were collected from the EPIC EHR system for the first 7 days following surgery or until hospital discharge if discharge occurred prior to POD 7. See Figure 1 for the study flow diagram.

FIGURE 1.

Study flow diagram

2.3 |. Description of outcomes

Following hospital discharge, patients were contacted monthly from 1 month postoperatively until 6 months postoperatively via telephone to complete the BPI. For the purposes of this analysis, the primary outcome of long-term pain following surgery was measured using BPI question 5 (BPIq5) pertaining to the mean pain intensity over the prior 24 hours. Cut points of None, Mild, Moderate, and Severe were established for ranges of 1 to 4 for Mild, 5 to 6 for Moderate, and 7 to 10 for Severe, as per the DVPRS descriptions, in accordance with prior ranges used for noncancer pain, and consistent with prior investigations using this methodology.(Buckenmaier et al., 2013; Polomano et al., 2016; Tighe et al., 2017; Woo et al., 2015) During patient interviews, clinical research coordinators specified that the BPI referred to pain specifically at the site of surgery rather than general body pain or remote pain locations.

2.4 |. Analysis

Analyses primarily focused on determining temporal profiles of acute postoperative pain through the use of SAX. The overarching goal of the analyses was to characterize how acute postoperative pain intensity changes over time on an assessment-to-assessment timeframe, and what, if any, predictive value exists in this information about the likelihood of developing PPP These characterizations, when stratified according to sociodemographic, behavioral, and PPP outcomes, aid in visualizing and quantifying those pain-transition motifs that characterize each strata.

We have previously detailed the methodology used for SAX across a range of biomedical domains, including postoperative pain, and the code used for these analyses are available in AppendixS1.(Nickerson et al., 2018; Tighe et al., 2017) We provide a brief overview of the steps below (Figure 2). For details about the statistical methods employed is examining these data, please refer to those prior publications.

FIGURE 2. Overview of SAX Methodology.

Panel A illustrates how raw time series data present complex signals for interpretation, with a random sample of five subjects highlighting several facets of time series patterns. Panels B and C illustrate how symbolic aggregate approximation (SAX) allows for the compression of time series data into vectorized format and then visual presentation using transition icons. Briefly, pain intensities are converted to symbols (here a,b,c,d on the y-axis of panel B), mildly similar to how sound waves in a song are converted to music notes; the aggregation of sequential pain intensity symbols then approximates the original time series using a process known as piecewise aggregate approximation (PAA). Each icon takes the form of a transition matrix demonstrating a two-step transition, with the rows representing the initial pain state with mild pain states at the upper rows and high pain states at the lower rows and the columns representing the subsequent pain state with mild pain states on the left and high pain states on the right. Icon cells in the left upper quadrant reflect transitions from low pain intensity states to low pain intensity states, those in the upper right reflect transitions from low pain intensity states to high pain intensity states, those in the lower left reflect transitions from high pain intensity states to low pain intensity states, and those in the right lower quadrant reflect transitions from high pain intensity states to high pain intensity states. The diagonal from upper left to lower right represents stable pain intensity, with the upper left representing stable, low-intensity pain. Transition along this diagonal would indicate stable pain with increasing intensity toward the lower right, which represents stable high-intensity pains. Axis tick marks represent quantiles (beta = 4 in this example) such that each pain intensity has an equiprobable change of being assigned a quantile.

Briefly, for each patient, we first organized the data to present an array of [subject ID, stratification variables, duration in minutes from end of Surgery to Pain Intensity Evaluation, Pain intensity] for each pain intensity observation. Pain intensities were then interpolated on a per-patient, per-minute basis to generate uniform time intervals between all pain scores analyzed (a necessary intervention for the algorithms to function). Saliency was determined for candidate temporal resolutions ranging from 20 to 60 minutes and revealed an optimal temporal resolution of 50 minutes, allowing the collapse of each 50-minute window into a single time epoch. We identified the optimal breakpoints in pain intensity values. Using the temporal resolution duration of 50 minutes, the distribution of the aggregate pain intensities within each 50-minute epoch were used to determine seven breakpoints. Following these temporal resolution and breakpoint determination steps, data were broken into a series of symbols, or “SAX words,” representing the transitions between various pain states. These were then arranged graphically into a square icon such that the upper left of each icon represented low-low transitions, upper right indicated low-high transitions, lower left high-low, lower right high-high, and the diagonal from upper left to lower right a stable state (Figure 2).

In the resulting graphic icons, the first (uppermost or leftmost) column/row represents zero pain, the second and third represent increasing levels of mild pain, the fourth and fifth represent increasing levels of moderate pain, and the sixth and seventh represent increasing levels of severe pain.

SAX representations using POD 0 to 7 acute pain intensities were generated according to outcomes for M1 and M6 time points. For the M6 time point, separate acute pain SAX representations were generated according to age groups, sex, surgical service, and pain catastrophizing.

SAX analyses offer a unique perspective on time series interpretation. The methods presented here improve on our prior work in this domain by extending outcome stratification to time points associated with subacute (M1) and chronic (M6) pain after surgery. Additionally, the use of an adaptive temporal resolution window ensured that the interpolation process did not lead to artifacts from overly small time period aggregations. The adaptive temporal resolution determination permitted a balanced tradeoff between longer windows, which can capture larger changes between adjacent frames, and shorter windows, which offer greater resolution to capture shorter trends.(Nickerson et al., 2018) In this particular set of experiments, the selection of beta = 7 provided the requisite resolution to demonstrate heterogeneity across the moderate transitions. Analyses were conducted using Python 3.4 and included the following libraries: Pandas, NumPy, SciPy, scikit-learn, joblib, multiprocessing, subprocess, matplotlib, seaborn, and tableone.(Hunter 2007; Jones et al., 2001; McKinney 2010; Pedregosa et al., 2011; Pollard et al., 2018; Van Der Walt et al., 2011; Varoquaux and Grisel 2009; Waskom et al., 2017)

3 |. RESULTS

3.1 |. Cohort characteristics

There were 364 unique subjects included in this cohort. The mean age was 59.25 years (SD 12.62) and 49.9% of subjects were female. Regarding surgical service, the most common services were colorectal surgery (n = 89, 24.7%), urology (n = 67, 18.6%), and thoracic surgery (n = 67, 18.6%; Table 1). Subject values for pain intensity were listed as “Patient Asleep” in 4,446 assessments out of 23,667 total assessments (18.8%) and were converted to missing. Prior to interpolation, there were 19,221 pain intensity scores across the cohort, or approximately 53 intensity scores per subject, with a mean of 4.21 and SD of 3.04. Following interpolation, there were 2.04×10⁶ pain intensity ratings, or approximately 5,610 per subject, with a mean of 3.85 and SD 2.76. Of all pain intensity assessments, 61.0% occurred on or before POD2.

TABLE 1.

Sociodemographics and postoperative pain outcome characteristics

Variable	Level	Missing	Count (Percent or Mean)
Subjects			364
Age Group	<40	3	30 (8.31)
	40–54		75 (20.78)
	55–70		173 (47.92)
	>70		83 (22.99)
Sex	Female	3	180 (49.86)
	Male		181 (50.14)
Race	Asian	3	2 (0.55)
	Black		32 (8.86)
	Hispanic		1 (0.28)
	Multiracial		1 (0.28)
	Other		15 (4.16)
	White		310 (85.87)
Ethnicity	Hispanic	3	14 (3.88)
	Not Hispanic		346 (95.84)
	Patient refused		1 (0.28)
Marital status	Divorced	3	28 (7.76)
	Life partner/significant other		4 (1.11)
	Married		211 (58.45)
	Other		1 (0.28)
	Separated		3 (0.83)
	Single		68 (18.84)
	Unknown		27 (7.48)
	Widowed		19 (5.26)
Surgical service	Colorectal surgery	3	89 (24.65)
	Neurosurgery		22 (6.09)
	Orthopaedics		60 (16.62)
	Pancreas & biliary surgery		55 (15.24)
	Thoracic cardiovascular surgery		67 (18.56)
	Urology		67 (18.56)
	Vascular surgery		1 (0.28)
Charlson Comorbidity Index		3	2.64 (3.01)
Postoperative Day Pain Intensity Assessment Count	0		2544 (13.24)
	1		5131 (27.7)
	2		3864 (20.1)
	3		2800 (14.57)
	4		1907 (9.92)
	5		1305 (67.89)
	6		928 (4.83)
	7		742 (3.86)
Preoperative Pain Catastrophizing Scale	0–29	133	202 (87.45)
	30–52		29 (12.55)
M1 Mean Pain Intensity	None		135 (41.28)
	Mild	37	129 (39.45)
	Moderate		45 (13.76)
	Severe		18 (5.5)
M6 Mean Pain Intensity	None		214 (74.56)
	Mild	77	53 (18.47)
	Moderate		12 (4.18)
	Severe		8 (2.79)
M1 Grouped Mean Pain Intensity	None or Mild		264 (80.73)
	Moderate or Severe	37	63 (19.27)
M6 Grouped Mean Pain Intensity	None or Mild		267 (93.03)
	Moderate or Severe	77	20 (6.97)

3.2 |. Acute postoperative pain trajectories and persistent postsurgical pain outcomes

For the entire cohort, the overall linear trajectory was one of decreasing intensity (Figure 3). Examination of simple linear acute pain trajectories suggests that patients reporting “None” for pain severity at 1 and 6 months following surgery consistently reported the lowest y-intercepts (4.02 M1, 4.24 M6, all for Pain Severity group “None”; Figure 3). Patients reporting pain of “None” had lower levels of pain intensity at all regressed time points than patients reporting pain of “Mild,” a pattern that remained consisted at M1 and M6 outcomes. Notably, the simple linear regressions suggested diverging patterns in slope and y-intercept for “Moderate” and “Severe” pain intensities at M1 and M6 time points. Additionally, the intersection between “Moderate” and “Severe” intensity outcomes across the M1 to M6 time points increased in time from operating room exit to pain assessment.

FIGURE 3. Linear Analyses of Postoperative Pain Time Series.

Linear trajectories of clinically collected acute postoperative pain intensity ratings on postoperative day 1 to 7 (a). The distribution of slopes generally differed across pain intensity ratings according to M1, M3, and M6 for those BPIq5 responses categorized as Moderate or Severe in intensity versus None or Mild. While the overall trajectory across the entire cohort was negative (b), examination across separate intensities yielded differing slopes and y-intercepts according to the BPIq5 intensity rating at M1(c) and M6 (d) time points. Notably the overall relative pattern of y-intercepts and slopes across different severity ratings were maintained across M1 and M6 outcomes Additionally, the overall intersection between moderate and severe intensity outcomes increased in time from operating room exit to pain assessment across the M1 to M6 time points.

3.3 |. SAX profiles for M1and M6 mean pain intensity outcomes

At 1 month postoperatively, patients who reported no pain had pain transitions during their acute phase (first 7 days postoperatively) predominantly in the low-to-low range (Figure 4). Patients who rated their pain as “Mild” at 1 month displayed a different pattern, with heterogeneous transitions, indicating that their acute pain was likely to have transitioned to many different intensities in the immediate postoperative period. A similar pattern was observed in patients who rated their pain “Moderate” at 1 month, albeit with somewhat more focus in the right lower quadrant (RLQ). In patients who rated their pain as “Severe” at 1 month, examination of their acute pain transitions showed a noteworthy absence along the diagonal from the upper left to lower right (this diagonal indicates stable pain of increasing severity, moving from stable mild pain in the upper left to stable severe pain in the lower right) except at the high end in the lower right. Instead, a preponderance of transitions to and from mostly moderate pain states was seen. Using the “None” pain intensity at M1 as reference, the cosine similarity, indicating the degree to which acute pain intensity transition patterns mirror each other when stratified by outcome group (i.e., how similar is the SAX icon for M1 “Mild” pain to “None”), was greatest for “Severe” (0.44) followed by “Moderate” (0.34) and “Mild” (0.33) pain intensity groups.

FIGURE 4. Symbolic Aggregate Approximation (SAX) Icons of Postoperative day 1 to 7 Pain Transitions Stratified by Mean Pain Intensity at 1 Month After Surgery.

Icon structures follow the template described in Figure 2. Each tick mark represents the quantile of the initial (y-axis) and following (x-axis) pain intensity.

When examining pain at 6 months postoperatively, among patients reporting no pain, the transitions in acute pain were largely heterogeneous, but, importantly, the RLQ was sparse. Patients reporting “Mild” pain at 6 months also showed significant heterogeneity in their acute pain transitions, with the left upper quadrant (representing transitions to and from minimal pain states) being notably sparse, while transitions to and from higher pain states were strongly represented. Among those reporting “Moderate” pain at 6 months, acute pain transitions were concentrated in the lower right half of the icon, representing transitions to and from moderate to severe pain during their immediate postoperative period. Patients reporting their pain as “Severe” at 6 months had acute pain transitions concentrated in the RLQ with a strip along the upper border and a strip along the left border of the icon. These represent severe-to-severe, moderate-to-severe, and moderate/severe-to-mild transitions, respectively (Figure 5). With the “None” pain intensity group at M6 as reference, the cosine similarity, indicating the degree to which acute pain intensities mirror outcome pain intensities, was greatest for the “Mild” (0.60) followed by “Severe” (0.49) and “Moderate” (0.36) groups.

FIGURE 5. Symbolic Aggregate Approximation (SAX) Icons of Postoperative day 1 to 7 Pain Transitions Stratified by Mean Pain Intensity at 6 Months After Surgery.

Icon structures follow the template described in Figure 2. Each tick mark represents the quantile of the initial (y-axis) and following (x-axis) pain intensity.

3.4 |. M6 outcome stratifications of SAX profiles by patient and surgical factors

When the pain outcomes at 6 months were separated based on sex, we found that in general, males displayed less heterogeneity than females. When examining patients reporting “None” or “Mild” pain at 6 months, females displayed acute pain transitions broadly scattered across all transition states while males predominantly had transitions to and from moderate/severe pain ratings with a paucity of stable pain states in the mild-to-moderate range. When examining patients reporting “Moderate” or “Severe” pain at 6 months, females had acute pain transitions concentrated in the RLQ, indicating transitions to and from moderate and severe pain, along with stable moderate and severe pain with few transitions to and from mild pain. Males, on the other hand, retained a pattern of acute pain transitions much more similar to those with no or mild pain at 6 months, predominantly transitions to and from moderate/severe pain ratings with a paucity of stable pain states in the mild-to-moderate range (Figure S1). The overall similarity between transition icons for M6 outcomes was greater for males (cosine similarity, indicating the degree to which acute pain intensities mirror outcome pain intensities, 0.71 between icons in male strata) than females (cosine similarity between icons in female strata 0.46)

When patient age was used as the separator, patients reporting no or minimal pain at 6 months who were under 40 years of age displayed pain transitions in the immediate aftermath of surgery that were concentrated in the upper left quadrant, indicating transitions to and from mild-to-moderate pain states, with a paucity of transitions to and from severe pain. In this same age group, patients who reported “Moderate” or “Severe” pain at 6 months essentially displayed the opposite pattern in their acute pain: a concentration in the RLQ, indicating transitions to and from moderate-to-severe pain states with an absence of transitions to and from mild pain.

Among patients 40 to 55 years of age, those who reported “None” or “Mild” pain at 6 months displayed a “shotgun” pattern of acute pain transitions, high numbers of transitions to and from most pain states with many holes in the icon, and an unpredictable pattern. Patients in the same age group who reported “Moderate” or “Severe” pain at 6 months still had significant heterogeneity in their acute pain transitions, but showed a greater concentration of transitions to and from moderate-to-severe pain in the week after surgery. Patients in the 55- to 70-year-old group who reported “None” or “Mild” pain at 6 months showed a pattern of acute pain transitions similar to that observed in the 40- to 54-year-old group with moderate-to severe pain at 6 months, but with increased heterogeneity, i.e., more concentration of transitions to and from moderate-to-severe pain in the week after surgery. Among those aged 55 to 70 reporting “Moderate” to “Severe” pain at 6 months, the icon displayed a RLQ predominance with a small density in the left lower quadrant, indicating transitions largely to and from moderate to severe, and transitions largely to mild from moderate to severe, respectively. Among those patients greater than 70 years of age, those with mild-to-moderate pain at 6 months displayed heterogeneous transitions in their acute pain with the notable exception of the left upper quadrant (LUQ), which would represent mild-to-mild transitions. Those in this age group who had severe pain at 6 months displayed focused clustering of their acute pain transitions in the RLQ, indicating that their pain transitions in the days immediately after surgery had been largely severe to severe (Figure S2). The greatest similarity in icons between M6 outcome categories was in the 55 to 70-year-old group (cosine similarity, indicating the degree to which acute pain intensities mirror outcome pain intensities, 0.61), followed by >70 years (cosine similarity 0.56), 40 to 54 years (cosine similarity 0.49), and <40 years (cosine similarity 0.32; Figure S2).

With respect to pain catastrophizing, we used a breakpoint score of 30 on the PCS to separate low catastrophizers from high catastrophizers. When pain outcomes at 6 months postoperatively were stratified based on PCS score, we found that among patients reporting “None” or “Mild” pain at 6 months, those with PCS scores under 30 showed a heterogeneous distribution of acute pain transitions with a slight emphasis in the RLQ. Overall, this suggests that their acute pain immediately after surgery was transitioning to and from all levels of intensity. Those with PCS scores over 30 showed a similar pattern with the LUQ-to-RLQ diagonal, which represents stable pain, being notably void. This indicates that their pain, too, was transitioning to and from all levels, but was rarely stable from one measurement to the next. Among those who rated their pain at 6 months as “Moderate” or “Severe,” patients with PCS scores under 30 displayed acute pain transitions immediately after surgery, which were predominantly transitions to and from severe pain states in the immediate aftermath of surgery, while those with PCS scores over 30 showed strong transitions along the right border of the icon, representing transitions predominantly to high pain states in the immediate aftermath of surgery with relatively few transitions to lower pain states (Figure S3). The PCS 0 to 29 group had greater cosine similarity between M6 outcome icons (0.55) than did the PCS 30 to 52 group (0.37).

When pain outcomes at 6 months after surgery were stratified based on the severity of patients’ preoperative pain, we found that among patients with no or mild pain at 6 months, both patients with no or mild pain preoperatively and those with moderate or severe pain preoperatively displayed heterogeneous acute pain transitions. However, those with moderate to severe pain preoperatively were less likely to show very low-level pain state transitions in the LUQ. Interestingly, patients with moderate to severe pain at 6 months who also had moderate to severe pain preoperatively displayed a similar pattern of acute pain transitions, both to and from all intensities of pain. However, among those with moderate to severe pain at 6 months who had no or mild pain preoperatively, acute pain transitions were concentrated in the RLQ, indicating persistently high levels of pain in the immediate aftermath of surgery (Figure S4).

Lastly, when pain outcomes at 6 months were separated based on the type of surgery, we found that across most types of surgery, similar patterns were displayed in acute pain: Patients who progressed to no or minimal pain displayed heterogeneous transitions to and from all pain intensities in the immediate aftermath of surgery, while those who progressed to moderate to severe pain at 6 months displayed much more focalization of their acute pain in higher pain transitions during the acute phase. The notable exceptions to this generalization included orthopaedic surgery, which displayed heterogeneous acute pain transitions among both those who went on to have no or mild pain at 6 months, as well as those who went on to have moderate to severe pain at 6 months; thoracic surgery, in which there was a vertical signal along the left border of the icon for patients with moderate to severe pain at 6 months, indicating a relatively high number of transitions to low pain states in the immediate postoperative period; and spine surgery, in which the general pattern described above appears to be reversed such that those with moderate to severe pain at 6 months displayed a great deal of heterogeneity in their acute pain transitions, while those with no to mild pain at 6 months displayed acute pain transitions more concentrated in the RLQ, suggesting more persistent, severe pain in the 7 days after surgery (Figure S5).

4 |. Discussion

As EHR systems become more common across the healthcare landscape, different information is becoming available to clinicians about their patients. Notably, all of this data is timestamped. While this may simply represent the fact that all computer functions are somehow tied to clock function, it also opens up the more likely possibility that when something happened is valuable information in addition to what happened. In the case of pain, the ability to look at pain scores over time and often convert those data into graphs within the EHR may have the potential to inform not only minute-to-minute actions to assist patients in acute pain, but also to shape choices that may impact long-term pain outcomes. Our results suggest that postoperative pain trajectories, based on clinical pain intensity ratings, in the 7 days after surgery are associated with differences in M1 and M6 pain severity outcomes. The overall trend in pain intensity decreased by 1 point over the first 7 days after surgery, a decrease notably smaller in magnitude than the 30% reduction in pain intensity representative of a clinically important difference.(Farrar et al., 2001)

In our analyses, SAX identified point-to-point heterogeneity in transition among pain states that were not fully captured, nor visualizable, by linear models. Icon morphologies for severe pain group were similar at M1 and M6 time points. On the contrary, icon morphology was different across the none and moderate severity pain groups. Patients reporting no pain at M6 had much more heterogenous patterns that included more transitions to and from moderate and severe pain intensity states than seen at the M1 time point. Patients reporting “Mild” pain at M1 through M6 all had elevated densities in the high-high transition regions, suggesting some level of discord with previous associations between increased acute pain intensity and PPP outcomes.(Gilron et al., 2017; Kampe et al., 2017; Katz et al., 1996; Kehlet et al., 2006) One potential explanation of this disagreement, based on intraday data, is that high-to-high pain intensity transitions may be more frequent but also more transient. Clinically, these results also suggest that over short time intervals, elevated pain intensity occurring across sequential measurements may not directly confer an increased risk of PPP, but rather that PPP risk is highest with persistently worsening or unrelenting acute pain.

Our data showed decreasing rates of moderate or severe pain intensity over the follow-up time course, ranging from 19.3% at M1 to 7.0% at M6. These findings align with work by Duale et al. showing decreases in pain intensity from M3 to M6 across a wide range of surgical procedures, although this particular cohort overlapped in surgical procedure types only in the thoracotomy subgroup.(Duale et al., 2014) While our data collection ended at M6, work by Fletcher at al. suggests continued decreases in pain intensity between M6 and 12 months after surgery.(Fletcher et al., 2015)

The stratification of transition icons by sex also identified increased RLQ transition density in females, suggesting more frequent transitions among elevated pain states, which agrees with prior literature showing increased overall postoperative pain intensity in females compared with males across a range of procedures in acute and PPP intensity.(Bartley and Fillingim 2013; Edgley et al., 2019; Fillingim 2017; Hah et al., 2019; Liu et al., 2012; Nandi et al., 2019; Tighe et al., 2014; Yang et al., 2019) The lower cosine similarity of transition icons between none/mild and moderate/severe M6 outcome groups in females compared with males also points to increased heterogeneity in acute pain transitions in females. This may represent greater transience of all pain states in females.

Stratification of transition icons at the M6 time point by age group was most notable for the increased transitions to moderate or severe pain with increasing age group in the none/mild M6 outcome and increasing consolidation to the RLQ with increasing age in the moderate/severe M6 outcome. While younger patients in this cohort with mixed transition patterns in POD1 to 7 may still progress to the moderate/severe M6 outcome category, older patients with moderate/severe M6 outcomes tended to have had acute pain score transitions more consistently in the high-to-high range. Prior reports on the relationship between age and PPP present conflicting results, with some data suggesting that increased age is protective for PPP and others showing no relationship, although this association may be partially masked by a peak in many pain conditions in middle age.(Cox et al., 2016; Edgley et al., 2019; Fillingim 2017; Hah et al., 2019; Hamood et al., 2018)

The rate of acute pain and PPP has been consistently shown to vary widely across types of surgery.(Gerbershagen et al., 2013; Katz and Seltzer; Kehlet et al., 2006; Perkins and Kehlet 2000) We observed significantly different patterns of distribution in transition icons when the M6 outcomes were stratified by surgery type. All surgery types had a RLQ focus for moderate/severe M6 outcomes, but orthopedic surgeries were unique given an increased transition density in the RLQ even for none/mild pain at M6. Spine and orthopedic surgeries, despite sharing a predominance of muscle and bone tissue involvement and significant rates of PPP, had different transition patterns in the none/mild M6 pain outcome category. On the other hand, surgeries involving the abdominal and pelvic organs (e.g., urology, colorectal, pancreas/biliary) were heterogenous in transition distribution for none/mild outcomes with consistent in presence in the RLQ, with frequent moves to/from the most severe intensities. Prior work by Gergershagen et al. highlighted increased pain intensities in orthopaedic, spine, and “minor” surgical procedures, but also noted the potential impact of regional anesthetics on the reported pain intensity for certain major surgical procedures.(Gerbershagen et al., 2013)

Prior works have repeatedly shown that pain catastrophizing is associated with an increased risk of both acute and chronic postsurgical pain.(Belfer et al., 2013; Edwards et al., 2013; Haren; Jacobsen and Butler 1996; Khan et al., 2011; Lewis et al., 2015; Pavlin et al. 2005,; Pinto et al., 2012; Quartana et al., 2009; Riddle et al., 2010; Ruscheweyh et al., 2011; Schreiber et al., 2013; Tuna et al., 2018) As expected, our results showed that low catastrophizers transitioned across a wide range of pain intensities for the none/mild M6 pain outcome, but remained in the severe/severe transition space for moderate/severe M6 outcomes. Clinically, this may suggest that a low catastrophizer may still experience episodes of severe acute postoperative pain, whereas low catastrophizers who do not enjoy at least some relief from severe acute postoperative pain may be at increased risk for PPP. On the other hand, for subjects with increased catastrophizing, the meaning of any given transition remains unclear given the degree of heterogeneity for any single transition, especially for transitions involving moderate pain intensity. Our results did suggest that high catastrophizers with none/mild M6 pain did not have transitions predominantly in the high-to-high range, suggesting that high-to-high transitions in catastrophizers could still indicate increased PPP risk.

In our cohort, 19.3% of subjects reported “Moderate” to “Severe” pain at M1 and 7.0% at M6. Lavand’homme summarized three key risk factors for PPP, including acute postoperative pain intensity, preoperative pain, and psychological distress.(Lavand’homme 2017) Our findings largely agree with this summary but offer additional insights into the intraday transitions of pain intensities that somewhat differ from prior findings using 24-hour aggregates. This agreement is reassuring given the potential for a temporal ecological fallacy in examining local versus more global changes of pain. Our results suggest that patients with elevated densities in the low-to-low transition space still have an association with an increased risk of PPP. It remains unclear if transient decreases in pain intensity reflect decrease nociceptive load, increased analgesic response, or other behavioral or physiologic measures mediating this response. Regardless, the data suggest that patients progressing to PPP are not necessarily consistent reporters of maximal pain intensities in the early postoperative period.

Our analyses encountered several limitations anticipated by the method of analyses. Our first limitation was in the management of “patient asleep” intensities. Given the large number of “patient asleep” values recorded, simply entering a value of “0” pain intensity would have led to a gross overrepresentation of “0” intensities in the distribution and likely further biased icon breakpoints in a downward fashion. Unfortunately, the vectors resulting from the SAX analyses are not readily usable in traditional statistical techniques given the resulting highly dimensional data. Our approach, while offering a guaranteed lower-bounded representation of the original time series and improved visual interpretation via the resulting transition icons, is limited to examining single-step transitions and thus likely loses information for consideration of relationships across longer time intervals and/or multi-step transitions. Additional methodologic advances are also necessary to incorporate qualitative data on each pain assessment alongside the intensity rating and to present these two dimensions together to provide appropriate clinical context for each intensity rating. For instance, Gilron et al. highlighted the important role of using rest versus movement-evoked metainformation in pain assessments in analyses of the association between persistent and postsurgical pain, context which unfortunately was not available in our clinical data collection schema.(Gilron et al., 2017)

Despite prior associations between acute pain and PPP intensity, the details of the underlying intraday acute pain intensity time series available to clinicians suggest that any given short-term pattern of pain transitions strongly depends on patient and procedural contexts. Here we have demonstrated that short-term patterns of pain transitions in higher pain intensity ranges remain commonly associated with no or mild pain at longer-term outcomes months after surgery, showing clearly that examining single point evaluations of pain, or even examining changes in pain from one evaluation to the next does not provide sufficient information to predict PPP outcomes. This work, along with other previous work, suggest the prognostic value of acute pain to PPP outcomes lies more in longer-term pain trajectories and examinations of transitions between multiple sequential pain states rather than single time point evaluations.

4.1 |. Conclusion

Clinically, these results suggest that individual pain-state transitions from one assessment to the next, even within patient or procedural strata associated with PPP, may not offer good predictive information with respect to the development of PPP. Further investigation of the time-pain relationship is clearly warranted. Our understanding of these interactions, how to interpret them, when and how to intervene, and what, if anything, they can tell us about the likelihood of pain-related problems that may be on the horizon remains in its infancy. Additionally, more work is necessary to incorporate such data into clinically available decision support tools, as well as to use complex time series data in more advanced statistical models.

Significance.

Symbolic aggregate approximations of clinically obtained, acute postoperative pain intraday time series identifies different motifs in patients suffering moderate to severe pain 6 months after surgery.