Data-driven identification of distinct pain drawing patterns and their association with clinical and psychological factors: a study of 21,123 patients with spinal pain

Supplemental Digital Content is Available in the Text.

Data-derived clusters corresponded to common, specific, and recognizable clinical presentations. Clusters of pain drawings were associated with activity limitation, pain intensity, and psychological distress.

Abstract

The variability in pain drawing styles and analysis methods has raised concerns about the reliability of pain drawings as a screening tool for nonpain symptoms. In this study, a data-driven approach to pain drawing analysis has been used to enhance the reliability. The aim was to identify distinct clusters of pain patterns by using latent class analysis (LCA) on 46 predefined anatomical areas of a freehand digital pain drawing. Clusters were described in the clinical domains of activity limitation, pain intensity, and psychological factors. A total of 21,123 individuals were included from 2 subgroups by primary pain complaint (low back pain (LBP) [n = 15,465]) or midback/neck pain (MBPNP) [n = 5658]). Five clusters were identified for the LBP subgroup: LBP and radiating pain (19.9%), radiating pain (25.8%), local LBP (24.8%), LBP and whole leg pain (18.7%), and widespread pain (10.8%). Four clusters were identified for the MBPNP subgroup: MBPNP bilateral posterior (19.9%), MBPNP unilateral posterior + anterior (23.6%), MBPNP unilateral posterior (45.4%), and widespread pain (11.1%). The clusters derived by LCA corresponded to common, specific, and recognizable clinical presentations. Statistically significant differences were found between these clusters in every self-reported health domain. Similarly, for both LBP and MBPNP, pain drawings involving more extensive pain areas were associated with higher activity limitation, more intense pain, and more psychological distress. This study presents a versatile data-driven approach for analyzing pain drawings to assist in managing spinal pain.

1. Introduction

Pain is a subjective experience, making it challenging to capture and quantify. However, a useful tool in clinical and research contexts is the pain drawing, which consists of an outline of the human body, onto which an individual can draw to indicate pain location and extent. Pain drawings can be easily administered in paper-based or digital formats and provide a quantifiable assessment of pain that can improve communication between healthcare professionals and patients.⁵¹

At face value, the pain drawing is simply an indication of the perceived anatomical location of pain, but it has been reported that pain drawings also hold information of diagnostic value in cases such as lumbar disk herniation, spinal stenosis, and nonspecific low back pain.^26,36,55 More commonly, it is claimed that pain drawings contain information about psychological factors such as depression, fear-avoidance beliefs, anxiety, social isolation, and catastrophization, all risk factors for a poor outcome.⁴⁰ Earlier studies suggest that specific pain drawing characteristics could indicate nonorganic symptoms reflecting psychological factors,^8,37,43,58 although this has been contested.^14,38,39,60

Pain drawings may be useful in diagnosis, but there is a risk of circular argumentation that specific pain drawing characteristics are linked to certain diagnoses, which could lead to confirmation bias in a clinical context. Studies have shown inconsistent results regarding the pain drawing's effectiveness as a screening tool for nonpain symptoms. Pain drawings, by their nature, are highly variable, but the analyses are also confounded by different methods and outcome measures from patients in different clinical settings.^6,16 For example, earlier pain drawing studies, such as those by Ransford⁴³ and Úden,⁵⁸ relied on subjective methods of interpretation and had questionable accuracy in detecting psychological distress in subacute and persistent low back pain.^1,4 Others have examined pain drawings in relation to diagnostic categories that were predefined based on clinical expertise.²⁶

A data-driven approach such as latent class analysis (LCA) has proven to have lower misclassification rates than traditional clustering approaches, making it a reliable method in spinal pain research.^10,15,21,54 In a study by Lacey et al.,²² around 13,000 individuals older than 50 years in North Staffordshire (United Kingdom) were analyzed, and 4 latent classes of pain patterns were identified. Among these classes, those reporting a “high number of painful sites” were more likely to also report cognitive disturbances and nonrestorative sleep. Similarly, Holden et al.¹⁷ identified distinct latent classes of pain patterns in nearly 3000 twelve- to nineteen-year-olds, where classes differed in sex, health-related quality of life, and sports participation.

This study had the overall objective of developing a robust approach to collecting and analyzing digital pain drawings routinely collected from hospital patients with spinal pain. We had 2 specific research aims of using LCA to (1) identify distinct latent classes, based solely on their pain pattern, and to describe these classes and (2) determine whether cluster membership is associated with activity limitation, pain intensity, and psychological factors.

2. Methods

2.1. Design and participants

This was an exploratory cross-sectional analysis of baseline data collected from an electronic clinical registry (SpineData).²⁰ Reporting follows the guidelines for STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) for cross-sectional studies.⁶¹ The Spine Centre of Southern Denmark is a large regional hospital department specializing in spinal pain syndromes. Referral criteria include persistent nonmalignant/noninfectious spinal pain, despite appropriate conservative treatment as per clinical guidelines in primary care.⁹ Every year since 2011, approximately 18,000 patients have been invited to complete an online questionnaire (SpineData) regarding their clinical presentation up to 2 days ahead of their initial consultation.²⁰ SpineData includes a digital pain drawing with 2 silhouette outlines of the human body onto which patients can draw their current perceived pain distribution by touch screen or computer mouse (Supplementary materials Figure S1, http://links.lww.com/PAIN/C46). Each patient is required to complete only one pain drawing per referral. If a patient experiences a new episode of pain that is not related to the primary pain in the first referral, they will be rereferred to the Spine Centre and asked to complete a new questionnaire relating to the new symptoms. Usually, patients who are rereferred have primary pain in their lower back during their first referral, and midback or neck pain during the next referral, or vice versa. SpineData's completion rate is around 80%. However, changes have been made to the questionnaire on several occasions because of newer versions of assessment tools, which are specified in section 2.4. Data between February 2019 and April 2021 were the most consistent regarding changes in the questionnaire and completion rate. Patients who completed a pain drawing in that period were included in this study. Patients were dichotomized based on their primary referral complaint as either low back pain (LBP) or midback/neck pain (MBPNP). If a patient had completed multiple questionnaires relating to the same primary complaint, they were removed from the corresponding subgroup.

2.2. Ethics

The Region of Southern Denmark was the data controller for this project (file no. 21/19378). Data processing in the project was regulated in accordance with the Danish Act on Research Ethics Review of Health Research Projects section 14, subsection 2. This section states that health research based solely on questionnaire surveys and registry data is exempt from the obligation to notify the ethics committees. We obtained approval for the use of data from hospital records for scientific purposes from the council of the Region of Southern Denmark (file no. 21/20523) in line with the Danish Health Care Act. The results presented in this article do not allow for the identification of subjects.

2.3. Data preprocessing

The freehand pain drawings were stored digitally in scalable vector graphics (SVG) format as a part of the questionnaire response. Each drawing could consist of any number of strokes (individual markings), and each stroke could consist of any number of points (Cartesian coordinates) connected by line segments along the stroke path. Pain drawings were converted from the raw SVG format into an array of stroke coordinates (Supplementary materials Figure S1, http://links.lww.com/PAIN/C46).

2.3.1. Stroke-polygons

The pain drawing coordinates for each stroke were converted into polygons using the MATLAB polyshape function⁵⁷ (The MathWorks, Inc. v.Natick, MA, United States. 2020a), which allowed for the calculation of the drawn pain area. The areas inside and outside the human body silhouette were estimated by the intersection of the pain drawing and silhouette. A polygon was created for each stroke by connecting raw data points in the same order as they were drawn and closing the polygon's first and last data points.³⁵ In this manner, any anatomical regions circumscribed by a pain drawing stroke, but not actually part of the stroke, would be included in the indicated area of pain. Illustrations of the polygon method can be seen in Supplementary materials Figure S1, http://links.lww.com/PAIN/C46.

2.3.2. Pain drawing characteristics

We identified the following pain drawing characteristics: number of anatomical regions, number of strokes and points, overlapping polygons, and area inside and outside the human body silhouette. The pain drawing polygons were analyzed in relation to 46 predefined anatomical regions based on the study by Margolis et al.,³⁰ which was not visible to patients when completing the pain drawing. The anatomical subdivision was superimposed on data only during analyses (Supplementary materials Figure S1, http://links.lww.com/PAIN/C46). The presence of pain in anatomically distinct areas was of relevance for our analyses, but the relevance of pain in the same anatomical areas on the left vs right side required careful consideration.²⁹ A priori, we decided that the important information was the presence of unilateral vs bilateral pain. In situations of unilateral pain, it was not considered important for analyses of pain drawing characteristics whether this was left- or right-sided. We reasoned that the presence of unilateral leg pain on the right side vs the left side represented specific details of unilateral leg pain and not a principally different characteristic of the pain drawing. For analyses, unilateral pain was mirrored to the left side of the body. A binary data set of pain drawing markings inside the 46 anatomical regions was then created and used. The area inside the human body silhouette was determined using MATLAB intersect function⁵⁶ (The MathWorks, Inc. v. 2020a). The number of pain sites was derived from the pain drawing as the sum of the 46 anatomical regions marked. The number of strokes, single points, and overlapping polygons was extracted from raw data. The areas inside and outside of the human body silhouette were calculated. Overlapping areas of polygons from different strokes were only counted once in the total area calculation.

2.4. Patient characteristics

SpineData includes several patient-reported outcomes regarding activity limitation, pain, and psychological factors. Patient age and sex were retrieved from their social security number, where age was calculated by subtracting the date of birth from the date of the initial consultation. Body mass index (BMI) was calculated from the patient-reported height and weight.

2.4.1. Activity limitation

Activity limitation was measured using Oswestry Disability Index (ODI, Danish version 2.1.a) for LBP and Neck Disability Index (NDI, 10-item version) for MBPNP. Both indexes consist of 10 items with a score from 0 to 5.^12,59 Activity limitation scores were calculated as the total score divided by the maximum total score times 100. In the presence of missing scores, a proportional recalculation was allowed for missing up to 2 items. This was performed by subtracting missing points from the maximum total score.¹¹ The score for activity limitation ranges from 0 to 100, with higher values representing greater disability.

2.4.2. Pain scores

Pain scores were reported using 0 to 10 numeric rating scales (NRSs) for current, worst, and typical pain over the previous 2 weeks.²⁸ Furthermore, NRS was reported for both regional (lower back or midback/neck) and extremity (upper or lower) pain, 6 items in total, with a maximum score of 60 points. Pain intensity was calculated as the sum of the 6 scores divided by the maximum total score times 100. Missing values were handled the same way described above for activity limitation.

2.4.3. Psychological factors

Psychological factors were measured using screening items with a 0 to 10 NRS, where values close to zero indicate less psychological distress. The psychological construct is outlined in the Supplementary materials Table S3, http://links.lww.com/PAIN/C46. The reasoning for the selection of these items is reported elsewhere.²⁰

2.5. Statistical methods

2.5.1. Model-based clustering and clinically relevant subgroups

Latent class analysis is an individual-oriented method for subgrouping.^23,33 The primary assumption behind LCA in this context was that a certain number of distinct spatial patterns of spinal pain existed, and individuals could be grouped into distinct classes based on their pain distribution, with each individual having the highest probability of belonging to one class. Latent class modeling aims to obtain the smallest number of classes that account for all the associations between the class-defining indicators (ie, the 46 predefined anatomical regions on a pain drawing). This approach contrasts with variable-based approaches (eg, regression and factor analysis) that do not typically account for individual variation.

2.5.2. Selecting the number of classes

Multiple criteria for determining the optimal number of classes have been proposed, but no single approach has been shown to be the most valid.^24,53,62 However, the most commonly used model selection techniques are Akaike information criterion (AIC) and Bayesian information criterion (BIC). In the case of studies with large sample sizes, the BIC is reported to be the most reliable for choosing the correct model.^24,34 Latent class analysis starts with a one-class model, assuming that all individuals have the same pain drawing, and then progressively adds an additional class for each successive model. Cluster analysis with a random start point can produce different results. To ensure global rather than local maxima, 100 iterations with 10 random starting points were used until the model converged, saturation was met, and maximum likelihood was found.⁵³ Posterior probabilities represent the certainty of an individual allocated to a specific class. Values close to one for the best-fitting class and zero for other classes indicate a high classification certainty. The posterior probability can be obtained for each individual allocated to the class for which this probability was the largest and plotted on a heat map. We decided that a mean posterior probability of no less than 90% was acceptable and would ensure the accuracy of latent class and reliability in the design of this study.⁶² A comparison between individuals with a posterior probability of 90% or more and posterior probability under 90% was inspected using the χ² test for categorical variables (sex) and Mann–Whitney U test for continuous variables. The AIC and BIC estimates and posterior probabilities were inspected for every class-model solution. Entropy can be another indicator of class certainty because this is a measure of separation between classes. Higher entropy denotes better class separation where values above 0.8 are acceptable. Although the entropy was not relied on in the final model to examine, it is good practice to report this.⁶² Latent class analysis was performed using the poLCA package in R (R Core Team, 2020, v. 1.6).²⁵ We expected the classes to represent pain distributions corresponding to common, specific, and recognizable clinical presentations.

2.5.3. Patient characteristics

Descriptive statistics are presented as numbers (n) and percentages (%) or mean values and standard deviations (SDs) or median and interquartile ranges (IQR) as appropriate. Q–Q plots were performed for visual inspection of normality, and the Levene test was used to test the homogeneity of variances.

2.5.4. Association with patient characteristics

Group differences in patient characteristics were assessed using the χ² test for categorical variables (sex) and one-way analysis of variance (ANOVA) for normally distributed continuous variables (activity limitation scores for MBPNP). The Kruskal–Wallis rank-sum test was used for variables with non-normally distributed data (age, BMI, activity limitation for LBP, pain, and psychological scores). Post hoc pairwise comparisons were conducted between LBP and MBPNP classes using appropriate test methods. A pairwise comparison between proportions was used for patient sex. The Tukey honest significant difference test was used for normally distributed data, and the Wilcoxon rank-sum test with Bonferroni correction was used for non-normally distributed data. The pairwise comparisons were followed by corresponding calculations of effect sizes, Wilcoxon effect size (r), or Cohen's D (d) as appropriate. The potential impact of missing data was considered negligible if the proportion of missing data was below 5%.¹⁸ The missing data were ignored in each analysis, and outliers interpreted as erroneous data were omitted. Analyses were two-tailed and considered statistically significant if the P-value < 0.05. Data analyses were conducted using R, for Windows, v. 4.0.3 with R-studio v. 1.1.4 software program, extended with relevant add-on packages.⁴²

3. Results

A total of 22,176 patients completed SpineData during the study period. However, patients with a blank pain drawing (4.75%) were excluded from analyses, leaving 21,123 individuals. Of these, 15,465 were classified with LBP (median age 56 years, 55.2% female, and median BMI of 27), and 5658 were classified as MBPNP (median age 51 years, 58.2% female, and median BMI of 26).

A 5-class model for LBP and a 4-class model for MBPNP were identified. Descriptive characteristics were inspected for each class (Tables 1–4). The classes labeled widespread pain had the highest percentage of females and were statistically significantly different from the other classes, except when comparing with LBP and whole leg pain and MBPNP unilateral posterior + anterior. The classes with the highest age on average were class 2 radiating pain (LBP) and class 4 MBPNP unilateral posterior. Statistical significant differences were found in age across all classes except when comparing LBP and radiating pain vs LBP and whole leg pain and MBPNP bilateral posterior vs widespread pain. Body mass index was very similar across all classes for both LBP and MBPNP. However, statistically significant differences were found, especially when comparing local LBP and radiating pain to widespread pain or MBPNP unilateral posterior to all other MBPNP classes. The proportion of missing values in each parameter was <5%, except for one item (fear of movement—“Physical activity might harm my back”) for MBPNP, with 24.1% missing data.

Table 1.

Description of individuals with low back pain stratified by latent class analysis, Spine Centre of Southern Denmark.

LCA-class characteristics	All individuals with LBP n = 15,465	Class 1: LBP and radiating pain n = 3073 (19.9%)	Class 2: radiating pain n = 3993 (25.8%)	Class 3: Local LBP n = 3829 (24.8%)	Class 4: LBP and whole leg pain n = 2896 (18.7%)	Class 5: Widespread pain n = 1674 (10.8%)	P and pairwise comparisons (P < 0.05)
Sex, females (%)	55.2%	54.1%	53.3%	50.5%	60.3%	64.1%	*P < 0.001 all comparisons except LBP and radiating pain (P adj. = 1), local LBP (P adj. = 0.12) vs radiating pain; LBP and whole leg pain vs widespread pain (P adj. = 0.13)
Age, y, median (IQR)	56 (43-68)	56 (43-69)	62 (50-72)	52 (38-66)	56 (44-68)	49 (37-60)	†P < 0.0001 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1)
BMI‡, median (IQR) Missings, n (%) Outliers, n (%)	27 (24-31) 234 (1.51%) 12 (0.08%)	27 (24-31)	27 (24-30)	27 (24-30)	27 (24-31)	27 (24-32)	†P < 0.05 all comparisons except LBP and whole leg pain vs LBP and radiating pain (P adj. = 0.85); radiating pain vs local LBP (P adj. = 0.11); LBP and radiating pain (P adj. = 1), local LBP (P adj. = 0.51), LBP and whole leg pain (P adj. = 0.58) vs widespread pain

Figures in Table 1: Class 1: LBP and radiating pain—Pain sites on the dorsum of the body. High pain intensity. Moderate to high activity limitation and psychological scores. Class 2: Radiating pain—Pain in one leg, frontal, and dorsal side of the body. Low to moderate activity limitation and pain intensity. Low psychological scores. Class 3: Local LBP—Localized pain in lower back and buttocks. Low activity limitation, pain intensity, and psychological scores. Class 4: LBP and whole leg pain—Bilateral low back and leg pain, frontal and dorsal. High activity limitation, pain intensity, and psychological scores. Class 5: Widespread pain—Multisite pain, whole body, frontal and dorsal. High activity limitation, pain intensity, and psychological scores.

^*χ² test.

^†Kruskal–Wallis test.

^‡Outliers = BMI < 7.5 or BMI > 105. In total, 12 individuals were outliers and were excluded from the analysis, as we interpreted these as error in data.

BMI, body mass index; IQR, interquartile range; LBP, low back pain; LCA, latent class analysis; P adj., P-value with Bonferroni correction.

Table 4.

Baseline characteristics of individuals with midback and neck pain stratified by latent class analysis, Spine Centre of Southern Denmark.

	All MBPNP individuals n = 5658	Class 1: MBPNP bilateral posterior n = 1124 (19.9%)	Class 2: MBPNP unilateral posterior + anterior n = 1334 (23.6%)	Class 3: MBPNP unilateral posterior n = 2570 (45.4%)	Class 4: Widespread pain n = 630 (11.1%)	P and pairwise comparisons (P <0.05)
Number of pain sites, median (IQR)	8 (5-12)	11 (8-13)	9 (7-12)	5 (3-7)	20 (18-25)	*P < 0.0001 all comparisons
Activity limitation†, 0-100 scale, mean (SD) Missings, n (%)	36.8 (17.0) 208 (3.7%)	38.7 (16.5)	36.9 (17.1)	33.5 (16.0)	47.0 (16.6)	§P < 0.05 all comparisons
Pain intensity‡, 0-100 scale, median (IQR) Missings, n (%)	60 (42-75) 44 (0.8%)	60 (40-75)	63 (47-77)	53 (37-72)	73 (60-83)	*P < 0.001 all comparisons
Current pain, 0-10 scale, median (IQR) Missings, n (%)	6 (4-8) 49 (0.9%)	6 (5-8)	6 (4-7)	6 (4-7)	7 (5-8)	*P < 0.0001 all comparisons except MBPNP unilateral posterior + anterior vs MBPNP unilateral posterior (P adj. = 0.11)
Typical pain, 0-10 scale, median (IQR) Missings, n (%)	7 (5-8) 62 (1.0%)	7 (6-8)	7 (5-8)	7 (5-8)	8 (6-8)	*P < 0.0001 all comparisons except MBPNP unilateral posterior + anterior vs MBPNP unilateral posterior (P adj. = 0.11)
Worst pain, 0-10 scale, median (IQR) Missings, n (%)	8 (7-9) 53 (0.9%)	8 (7-9)	8 (7-9)	8 (6-9)	9 (8-10)	*P < 0.01 all Comparisons
Anxiety, “Do you feel anxious?” 0-10 scale, median (IQR) Missings, n (%)	2 (0-5) 133 (2.4%)	3 (0-6)	3 (0-5)	2 (0-5)	4 (1-7)	*P < 0.05 all comparisons except MBPNP bilateral posterior vs MBPNP unilateral posterior + anterior (P adj. = 0.84)
Social isolation, “Do you feel socially isolated?” 0-10 scale, median (IQR) Missings, n (%)	0 (0-2) 131 (2.3%)	0 (0-2)	0 (0-2)	0 (0-1)	0 (0-3)	*P < 0.001 all comparisons except MBPNP bilateral posterior (P adj. = 0.27), MBPNP unilateral posterior (P adj. = 0.30) vs MBPNP unilateral posterior + anterior
Catastrophization, “When I feel pain, it is terrible and I feel it is never going to get any better”, 0-10 scale, median (IQR) Missings, n (%)	5 (2-8) 149 (2.6%)	5 (3-8)	5 (2-7)	5 (2-7)	6 (4-8)	*P < 0.01 all comparisons except MBPNP unilateral posterior vs MBPNP unilateral posterior + anterior (P adj. = 0.16)
Catastrophization, “When I feel pain, I feel I cannot stand it anymore,” 0-10 scale, median (IQR) Missings, n (%)	4 (1-7) 140 (2.5%)	5 (2-7)	4 (1-7)	3 (1-6)	5 (3-8)	*P < 0.05 all comparisons
Depression, “During the past month, have you often been feeling down, depressed, or hopeless,” 0-10 scale, median (IQR) Missings, n (%)	2 (0-6) 151 (2.7%)	3 (1-7)	2 (0-6)	2 (0-5)	5 (1-7)	*P < 0.01 all comparisons
Depression, “During the past month, have you been bothered by reduced interest or pleasure in doing things?” 0-10 scale, median (IQR) Missings, n (%)	3 (0-7) 164 (2.9%)	4 (1-7)	3 (0-7)	2 (0-6)	5 (2-8)	*P < 0.05 all comparisons
Fear of movement, “Physical activity might harm my back," 0-10 scale, median (IQR) Missings, n (%)	5 (1-7) 1361 (24.1%)	5 (2-8) 280 (24.9%)	3 (1-6) 318 (23.8%)	5 (1-7) 623 (24.2%)	5 (2-8) 140 (22.2%)	*P < 0.01 all comparisons except MBPNP unilateral posterior + anterior vs MBPNP unilateral posterior (P adj. = 1); MBPNP bilateral posterior vs widespread pain (P adj. = 0.14)
Fear of movement, “I should avoid physical activities which (might) make my pain worse,” 0-10 scale, median (IQR) Missings, n (%)	5 (1-7) 227 (4.0%)	5 (1-8)	5 (1-7)	4 (1-7)	5 (2-8)	*P < 0.05 all comparisons except MBPNP bilateral posterior (P adj. = 0.32), MBPNP unilateral posterior (P adj. = 1) vs MBPNP unilateral posterior + anterior
Self-perceived risk of persistent pain, “In your view, how large is the risk that your current pain may become persistent?” 0-10 scale, median (IQR) Missings, n (%)	7 (5-9) 172 (3.0%)	8 (5-10)	7 (5-9)	7 (5-9)	8 (6-10)	*P < 0.001 all comparisons except MBPNP unilateral posterior + anterior vs MBPNP unilateral posterior (P adj. = 1)

^*Kruskal–Wallis test.

^†Activity limitation was calculated from 10 NDI questions.

^‡Pain intensity was calculated, with the same method used for NDI scores (†). Pain intensity was calculated as a sum of 6 scores—midback/neck pain and arm pain in 3 scenarios (current, typical, and worst) divided by the maximum total score times 100.

^§One-way ANOVA

IQR, interquartile range; MBPNP, midback/neck pain; P adj., P-value with Bonferroni correction.

Table 2.

Description of individuals with midback and neck pain stratified by latent class analysis, Spine Centre of Southern Denmark.

LCA-class characteristics	All MBPNP individuals n = 5658	Class 1: MBPNP bilateral posterior n = 1124 (19.9%)	Class 2: MBPNP unilateral posterior + anterior n = 1334 (23.6%)	Class 3: MBPNP unilateral posterior n = 2570 (45.4%)	Class 4: Widespread pain n = 630 (11.1%)	P and pairwise comparisons (P < 0.05)
Sex, females (%)	58.2%	57.4%	60.8%	55.3%	66.5%	*P < 0.01 all comparisons except MBPNP unilateral posterior + anterior (P adj. = 0.57), MBPNP unilateral posterior (P adj. = 1) vs MBPNP bilateral posterior; MBPNP unilateral posterior + anterior vs widespread pain (P adj. = 0.1)
Age, y, median (IQR)	51 (41-60)	48 (37-59)	51 (42-59)	53 (42.2-62)	48 (38-55)	†P < 0.001 all comparisons except MBPNP bilateral posterior vs widespread pain (P adj. = 1)
BMI‡, median (IQR) Missings, n (%) Outliers, n (%)	26 (24-30) 71 (1.3%) 1 (0.02%)	27 (24-30)	27 (24-30)	26 (23-29)	27 (24-31)	†P < 0.01 all comparisons except MBPNP unilateral posterior + anterior (P adj. = 1), widespread pain (P adj. = 1) vs MBPNP bilateral posterior; MBPNP unilateral posterior + anterior vs widespread pain (P adj. = 1)

Figures in Table 2: Class 1: MBPNP bilateral posterior—Pain sites on the dorsum of the body. Average pain intensity. Moderate to high activity limitation and psychological scores. Class 2: MBPNP unilateral posterior + anterior—Frontal and dorsal pain on upper extremities. Demographics and baseline scores are similar to subpopulation mean. Class 3: MBPNP unilateral posterior—Localized dorsal pain in upper extremities. Low activity limitation, pain intensity, and psychological scores. Class 4: Widespread pain—Multisite pain, whole body, frontal and dorsal. High activity limitation, pain intensity, and psychological scores.

^*χ² test.

^†Kruskal–Wallis test.

^‡Outliers = BMI < 7.5 or BMI > 105. In total, one individual was outlier and was excluded from the analysis, as we interpreted these as error in data.

BMI, body mass index; IQR, interquartile range; LCA, latent class analysis; MBPNP, midback/neck pain; P adj., P-value with Bonferroni correction.

Table 3.

Baseline characteristics of individuals with low back pain stratified by latent class analysis, Spine Centre of Southern Denmark.

	All individuals with LBP n = 15,465	Class 1: LBP and radiating pain n = 3073 (19.9%)	Class 2: Radiating pain n = 3993 (25.8%)	Class 3: Local LBP n = 3829 (24.8%)	Class 4: LBP and whole leg pain n = 2896 (18.7%)	Class 5: Widespread pain n = 1674 (10.8%)	P and pairwise comparisons (P < 0.05)
Number of pain sites, median (IQR)	6 (4-10)	7 (6-9)	3 (2-5)	5 (4-6)	11 (9-13)	15 (12-20)	*P < 0.0001 all comparisons
Activity limitation†, 0-100 scale, median (IQR) Missings, n (%)	34 (22-48) 571 (3.7%)	36 (26-50)	32 (20-46)	30 (20-42)	38 (26-52)	38 (26-50)	*P < 0.01 all comparisons except LBP and radiating pain (P adj. = 0.65), LBP and whole leg pain (P adj. = 1) vs widespread pain
Pain intensity‡, 0-100 scale, median (IQR) Missings, n (%)	60 (42-77) 141 (0.9%)	68 (53-80)	57 (40-74)	45 (33-63)	68 (53-80)	67 (50-80)	*P < 0.05 all comparisons except LBP and radiating pain (P adj. = 1), widespread pain (P adj. = 0.11) vs LBP and whole leg pain; LBP and radiating pain vs widespread pain (P adj. = 0.50)
Current pain, 0-10 scale, median (IQR) Missings, n (%)	6 (4-7) 203 (1.3%)	6 (4-8)	5 (2-7)	6 (4-7)	6 (4-8)	7 (5-8)	*P < 0.01 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 0.09)
Typical pain, 0-10 scale, median (IQR) Missings, n (%)	7 (5-8) 259 (1.7%)	7 (5-8)	6 (4-8)	7 (5-8)	7 (5-8)	7 (6-8)	*P < 0.05 all comparisons
Worst pain, 0-10 scale, median (IQR) Missings, n (%)	8 (7-9) 226 (1.5%)	8 (7-10)	8 (5-9)	8 (7-9)	8 (7-10)	9 (8-10)	*P < 0.001 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1)
Anxiety, “Do you feel anxious?” 0-10 scale, median (IQR) Missings, n (%)	4 (1-7) 432 (2.8%)	5 (2-7)	4 (1-6)	4 (1-7)	5 (2-7)	5 (2-8)	*P < 0.01 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1)
Social isolation, “Do you feel socially isolated?” 0-10 scale, median (IQR) Missings, n (%)	0 (0-2) 426 (2.8%)	0 (0-3)	0 (0-2)	0 (0-2)	0 (0-3)	1 (0-4)	*P < 0.001 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1); radiating pain vs local LBP (P adj. = 0.85)
Catastrophization, “When I feel pain, it is terrible and I feel it is never going to get any better,” 0-10 scale, median (IQR) Missings, n (%)	5 (3-8) 486 (3.1%)	6 (3-8)	5 (2-8)	5 (3-8)	6 (3-8)	6 (3-8)	*P < 0.01 all comparisons except LBP and whole leg pain (P adj. = 1), widespread pain (P adj. = 0.06) vs LBP and radiating pain; radiating pain vs local LBP (P adj. = 0.31)
Catastrophization, “When I feel pain, I feel I cannot stand it anymore,” 0-10 scale, median (IQR) Missings, n (%)	5 (2-8) 470 (3.0%)	5 (2-8)	4 (1-7)	4 (1-7)	5 (2-8)	5 (2-8)	*P < 0.001 all comparisons except LBP and whole leg pain (P adj. = 1), widespread pain (P adj. = 0.09) vs LBP and radiating pain; radiating pain vs local LBP (P adj. = 1); LBP and whole leg pain vs widespread pain (P adj. = 0.21)
Depression, “During the past month, have you often been feeling down, depressed, or hopeless,” 0-10 scale, median (IQR) Missings, n (%)	4 (1-7) 470 (3.0%)	4 (1-7)	3 (0-6)	3 (0-6)	4 (1-7)	5 (2-8)	*P < 0.0001 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1) radiating pain vs local LBP (P adj. = 1)
Depression, “During the past month, have you been bothered by reduced interest or pleasure in doing things?”—0-10 scale, median (IQR) Missings, n (%)	5 (2-8) 535 (3.5%)	5 (2-8)	5 (1-7)	5 (1-7)	5 (2-8)	6 (2-8)	*P < 0.05 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1); radiating pain vs local LBP (P adj. = 1); LBP and whole leg pain vs widespread pain (P adj. = 0.11)
Fear of movement, “Physical activity might harm my back,” 0-10 scale, median (IQR) Missings, n (%)	5 (1-7) 721 (4.7%)	5 (2-8)	3 (1-6)	5 (1-7)	5 (1-7)	5 (2-8)	*P < 0.001 all comparisons except local LBP (P adj. = 0.15), widespread pain (P adj. = 1) vs LBP and radiating pain; LBP and whole leg pain (P adj. = 0.08), widespread pain (P adj. = 0.06) vs local LBP
Fear of movement, “I should avoid physical activities which (might) make my pain worse,” 0-10 scale, median (IQR) Missings, n (%)	5 (2-8) 718 (4.6%)	5 (2-8)	5 (1-8)	5 (2-8)	5 (2-8)	5 (2-8)	*P < 0.01 all comparisons except local LBP (P adj. = 0.14), widespread pain (P adj. = 1) vs LBP and radiating pain; LBP and whole leg pain (P adj. = 1), widespread pain (P adj. = 0.07) vs local LBP
Self-perceived risk of persistent pain, “In your view, how large is the risk that your current pain may become persistent?” 0-10 scale, median (IQR) Missings, n (%)	8 (5-9) 589 (3.8%)	8 (5-10)	7 (5-9)	7 (5-9)	8 (5-9)	8 (6-10)	*P < 0.05 all comparisons except LBP and radiating pain vs LBP and whole leg pain (P adj. = 1)

^*Kruskal–Wallis test.

^†Activity limitation was calculated from 10 ODI questions.

^‡Pain intensity was calculated, with the same method used for ODI scores (†). Pain intensity was calculated as a sum of 6 scores—low back pain and leg pain in 3 scenarios (current, typical, and worst) divided by the maximum total score times 100.

IQR, interquartile range; LBP, low back pain; P adj., P-value with Bonferroni correction.

3.1. Latent class analysis model selection

Results from LCA and model fit estimates suggested an optimal number of classes much higher than expected (Supplementary materials Figure S2, http://links.lww.com/PAIN/C46). As the models became more complex, the improvement in fit was, to some degree, driven by the large sample size. The BIC kept lowering beyond a reasonable number of classes. Thus, we decided not to increase the number of classes with every repetition whenever the size of one class became smaller than 10% (Supplementary materials Table S4, http://links.lww.com/PAIN/C46). Although BIC decreased for every class added, posterior probabilities stayed within the acceptable range (≥90%). Posterior probabilities for a four- to six-class-model solution were inspected for LBP (Supplementary materials Table S5, http://links.lww.com/PAIN/C46), and posterior probabilities for a three- to five-class-model solution were inspected for MBPNP (Supplementary materials Table S6, http://links.lww.com/PAIN/C46). The five-class model for LBP and the four-class model for MBPNP were deemed most relevant to inspect, primarily based on class size and posterior probabilities, and secondarily on visual inspection of aggregated pain drawings as being clinically distinct. The classes derived from LCA had a mean posterior probability between 92% and 97%, indicating high confidence in individuals allocated to their respective classes. The lowest probability of any individual belonging to a class was 38% (LBP) and 39% (MBPNP) (Supplementary materials Table S7 + S8, http://links.lww.com/PAIN/C46). However, 79.77% of LBP and 90.05% of MBPNP individual's posterior probability of belonging to their allocated class were 90% or higher. The remaining 20.23% for LBP and 9.95% for MBPNP were still allocated to the class where posterior probability was highest, although with higher uncertainty. Statistical differences in 2 psychological items (Catastrophization—“When I feel pain, I feel I cannot stand it anymore” and self-perceived risk of persistent pain) were found when comparing individuals with LBP with a posterior probability of 90% or more and those with a posterior probability under 90%, see Table S9, http://links.lww.com/PAIN/C46. In the same comparisons for the MBPNP individuals, statistical differences were found in age (51 years [IQR: 41-60] vs 49 years [IQR: 38-59]), activity limitation, and 5 psychological items, see Table S10, http://links.lww.com/PAIN/C46.

3.2. Latent class analysis classes

The LCA classes were visualized by heat maps of the probability of an individual having membership in a specific class, where the darker the color, the higher the probability of indicating pain in that anatomical region, see Figure 1. The classes were labeled according to clinically meaningful pain distribution patterns.

Figure 1.

This figure shows the aggregated pain drawings in classes derived from LCA for LBP (upper row; 5-class-model solution) and MBPNP (lower row; 4-class-model solution). The frequency of pain marked in each of the 46 anatomical regions is presented with a red scale heat map (0%-100%). The darker the color of red, the more individuals have indicated pain in that anatomical region. Descriptions of the main characteristics of each class are provided under the respective aggregated pain drawing. When inspecting the aggregated pain drawings, the reader should keep in mind that any difference in shading between the left and right side represents a difference in the frequency of unilateral vs bilateral pain indications, respectively, as any unilateral right-sided pain indication has been mirrored to the left side. LBP, low back pain; LCA, latent class analysis; MBPNP, midback pain/neck pain.

3.2.1. Latent class analysis classes for low back pain

Class 1 was labeled as LBP and radiating pain (19.9% [n = 3073]) and was characterized by multiple pain sites limited to the dorsum of the body with high pain intensity and psychological scores. The largest class was class 2, labeled as radiating pain (25.8% [n = 3993]). Class 2 was characterized by pain extending into one leg, reporting the highest mean age, the smallest number of pain sites, moderate activity limitation and pain intensity, and the lowest mean scores on psychological factors. Class 3 was labeled as local LBP (24.8% [n = 3829]) and was characterized by the lowest activity limitation, pain intensity, and psychological scores on average. Class 4 was labeled as LBP and whole leg pain (18.7% [n = 2896]) and was characterized by LBP and primarily unilateral leg pain (both frontal and dorsal), reporting high activity limitation, pain intensity, and psychological scores. The smallest class was class 5, labeled as widespread pain (10.8% [n = 1674]). Class 5 was characterized by high activity limitation and pain intensity while having the highest scores on average in all psychological factors.

3.2.2. Latent class analysis classes for midback/neck pain

Class 1 was labeled as MBPNP bilateral posterior (19.9% [n = 1124]). Class 1 shared the same characteristics as class 1 for LBP but with primary pain in the upper-body regions. Class 2 was labeled as MBPNP unilateral posterior + anterior (23.6% [n = 1334]) and was characterized by pain in both the front and dorsum of the upper body extending into the upper extremities. Class 2 reported activity limitation, pain intensity, and psychological scores closest to the subpopulation mean. The largest class was class 3, labeled as MBPNP unilateral posterior (45.4% [n = 2570]), and had more localized pain restricted to the dorsum of the body compared with class 2. Class 3 was characterized by the highest mean age and the lowest average in activity limitation, pain intensity, and psychological scores. Class 4 was labeled as widespread pain (11.1% [n = 630]) and shared the same characteristics as class 5 for LBP but with primary pain in the upper-body regions.

3.3. Association between latent class analysis classes

Overall, individuals in classes with more extended pain sites presented higher scores on all clinical characteristics. Classes labeled as widespread pain had statistically significant higher scores on psychological factors compared with other classes. Notably for both LBP and MBPNP, classes with extensive pain areas had significantly higher activity limitation, pain intensity, and psychological distress than classes with more localized pain.

3.3.1. Low back pain classes

The activity limitation scores in class 2 radiating pain and class 3 local LBP were significantly lower (P < 0.001) compared with class 1 LBP and radiating pain, class 4 LBP and whole leg pain, and class 5 widespread pain. However, the effect sizes were small. The pain intensity was significantly higher for widespread pain (median = 67) with small and moderate effect sizes when compared with radiating pain (median = 57, P < 0.05, r = 0.165) and local LBP (median = 45, P < 0.05, r = 0.346). Class 3 local LBP reported significantly lower pain intensity scores and showed moderate effect sizes when compared with class 1 LBP and radiating pain (median = 68, P < 0.05, r = 0.416) or class 4 LBP and whole leg pain (median = 68, P < 0.05, r = 0.403). When assessing all psychological factors, there was a clear pattern of the classes with more extensive pain areas (class 1 LBP and radiating pain, class 4 LBP and whole leg pain, and class 5 widespread pain) having higher median scores compared with the classes with more localized pain in a specific area (class 2 radiating pain and class 3 local LBP). Statistical significant differences only appeared when comparing between these groups, but not within the groups. However, only class 2 radiating pain had significantly lower scores in the “fear of movement” parameters, compared with the other classes. Although statistically significant differences were found in the psychological factors, the effects were small.

3.3.2. Midback/neck pain classes

The activity limitation scores were significantly higher for class 4 widespread pain (mean = 47.0, SD = 16.6) with moderate to large effect sizes when compared with class 1 MBPNP bilateral posterior (mean = 38.7, SD =16.5, P < 0.05, d = 0.502), class 2 MBPNP unilateral posterior + anterior (mean = 36.9, SD = 17.1, P < 0.05, d = 0.601), and class 3 MBPNP unilateral posterior (mean = 33.5, SD = 16.0, P < 0.05, d = 0.832). The pain intensity was significantly higher for class 4 widespread pain (median = 73) with small to moderate effect sizes when compared with class 1 MBPNP bilateral posterior (median = 60, P < 0.001, r = 0.286), class 2 MBPNP unilateral posterior + anterior (median = 63, P < 0.001, r = 0.210) and class 3 MBPNP unilateral posterior (median = 53, P < 0.001, r = 0.302). When assessing the psychological factors, the classes MBPNP bilateral posterior and widespread pain were significantly different from classes MBPNP unilateral posterior and MBPNP unilateral posterior + anterior, except for the anxiety and depression parameters. However, the psychological factors only had small effects on the MBPNP classes.

3.4. Pain drawing characteristics

Individuals in LCA classes with a high number of pain sites had larger calculated areas inside and outside the pain drawing. Specifically, the classes labeled widespread pain (for LBP and MBPNP) had larger areas outside the pain drawing with a factor of 1:18 and 1:10 compared with all LBP or MBPNP individuals, respectively (Supplementary materials Table S1 + S2, http://links.lww.com/PAIN/C46).

4. Discussion

The digitalization of pain drawings has enabled new methods of quantifying pain drawings, but only a few studies have used a data-driven approach to investigate subgroups of pain drawings from individuals with persistent spinal pain.^2,3,22,47 Based on self-reported pain drawings, we identified a five- and four-latent-class-model solution for LBP and MBPNP, respectively. We found that classes with more extensive pain patterns were highly associated with more activity limitation, more intense pain, and more psychological distress. Importantly, the classes derived through data-driven analyses represent pain distributions corresponding to common, specific, and recognizable clinical presentations that provide valuable information beyond the pain extent and location. Besides aiding in prognosis research, minimizing misclassification, and assisting with recognizing specific spinal pain patterns, pain drawing classes could potentially inform inclusion criteria for clinical trials evaluating the effectiveness of tailored interventions.

Clustering is substantially influenced by variable selection, and no gold standard for choosing the optimal number of classes in LCA exists. Therefore, we focused on the practical usefulness of LCA, relying on the number of individuals in each class and the parsimony of the LCA model solutions.³¹ We are confident that our LCA models accurately classified each individual appropriately because the mean posterior probability was greater than 90% for all LBP and MBPNP classes.

4.1. Clinical implications

The study was conducted on a hospital population, but we suspect similar conclusions apply in primary care¹³—after all, the majority were referred from primary care. Pain drawings are used extensively as a clinical tool, and as far back as 1976, they have been used as an aid in the psychological evaluation of patients with LBP.⁴³ A systematic review conducted in 2018 concluded that there is no definitive answer to the potential correlation between pain drawings and psychological factors.⁴⁵ Our study provides robust support to the hypothesis that pain drawings do contain information related to nonpain symptoms in an LBP and MBPNP population such as depression and anxiety.

Studies have shown that regardless of the location of pain, individuals with pain in more than one area report significantly higher disability and reduced overall health.^19,32,46,49 Likewise, recent studies have shown a strong linear association between the number of pain sites and higher psychological distress.^13,64 This implies that defining specific anatomical regions could affect the optimal number of clusters and their description. In this study, we used 46 anatomical regions based on Margolis et al.,³⁰ which has demonstrated high reliability in previous research.¹⁶ The study by Alter et al.² included pain drawings from 21,685 patients seen at a multidisciplinary pain care unit. They identified 9 clusters of pain drawings, most of which were related to spinal pain, based on 74 anatomical regions and hierarchical clustering. They found that widespread pain compared with localized pain was associated with higher activity limitation, pain intensity, anxiety, and depression scores, which are in agreement with our findings. The study by Lacey et al.²² identified a four-class model solution using LCA on pain drawings divided into 16 anatomical regions from primary care patients aged 50 years or older (N = 12,408). The study by Holden et al.¹⁷ used the same method, but with 12 anatomical regions on pain drawing from adolescents in the age of 12 to 19 years (N = 2953). In both studies, the classes were defined based on the median number of pain sites and the likelihood of reporting pain in different body regions such as the lower back. Elderly primary care patients with a large number of pain sites had significantly higher levels of anxiety and depression. Adolescents with a large number of pain sites had significantly lower health-related quality of life. Although the methods of pain drawing delineation and the numbers of clusters were different from our study, we reached similar conclusions. It is worth noting that certain pain patterns observed in our study were similar to those found in other studies, indicating the possibility of a universal classification system. However, since we did not have access to clinical diagnoses in our study, it is crucial for future studies to explore this aspect to achieve a universal classification system.

Our findings should not be interpreted to mean that pain drawings are of diagnostic utility.^27,41,44 For example, membership in class 1 LBP and radiating pain and class 2 radiating pain should raise clinical suspicion of radicular pain, which is typically seen with a disk herniation, but taken alone, is not sufficient for such a diagnosis. Similarly, membership in the widespread pain class (class 4 for MBPNP and class 5 for LBP) should raise clinical awareness of the possibility of generalized hyperalgesia, psychological distress, and/or nonorganic pain components, but is also not diagnostic. Future research is needed to investigate the diagnostic and prognostic value of pain patterns in these clinical populations.

4.2. Limitations

The development and use of pain drawings have been evolving for decades.^48,52 However, the delineation of the anatomical regions (eg, body maps or pain drawings) differs.^5,50 Despite recent advancements in pain drawing analysis, some individuals experience more complex patterns of pain that cannot be easily captured by simplified pain drawing variables or a single body-site code.^7,63 The lack of standardization of the human body outline and pain drawing quantification methods may hinder comparison between studies (eg, when calculating a geographic area or counting anatomical regions). Establishing a universal standard for automated pain drawing quantification is challenging because of the influence of the drawing style. For instance, Bryner et al.⁵ found that pain area is overestimated when using grid-based assessment and underestimated when using a computer-based pixel system for small areas on a pain drawing. To improve pain representation, we used the simple polygon method³⁵ and assessed markings within 46 anatomical regions.³⁰ This approach is useful when a patient has drawn a circle around an anatomical area and an enclosed region is circumscribed, but not actually marked, thus potentially not counted. Conversely, nonpainful anatomical areas around a painful area may be wrongly included when a patient encircles the painful area. Moreover, relying on the actual marking rather than anatomical regions is technically more challenging.

It is important to note that our implementation of the pain drawing may differ from other studies, which may have implications for the generalizability of our findings. However, a study by Lacey et al.²² suggests that a blank pain drawing, such as the one used in our study, is more appropriate for patients with persistent spinal pain who experience longer pain duration, more disability, and worse severity as compared to acute spinal pain. In any case, when using a digital pain drawing, there are technical issues to attend to, and some patients may find it difficult to interface with a touchscreen or use a computer mouse to draw on-screen.

Although mean posterior probabilities were high, finding the optimal LCA model was not straightforward as we could not rely on AIC or BIC. To ensure that our model represents clinically relevant subgroups in the larger population and not just individual outliers, we opted for solutions where each class contained at least 10% of the study population. These were pragmatic choices, and it is possible that other approaches would have led to different outcomes.²⁴

It was observed that approximately 20% (LBP) and 10% (MBPNP) had a posterior probability below 90% (Table S9+S10, http://links.lww.com/PAIN/C46). Although these individuals were assigned to the class with the highest posterior probability, uncertainty in classification might persist. It was observed that only catastrophization and risk of persistent pain were statistically significant for individuals with LBP. This indicates that individuals with LBP who were allocated to a class with higher certainty reported higher scores in these psychological items. Statistical differences in age, activity limitation, catastrophization, depression, fear of movement, and risk of persistent pain were found in the MBPNP subgroup. Individuals allocated to classes with higher certainty were older but reported lower activity limitation and psychological scores.

Patient-reported outcomes are prone to missing data, but we did not investigate whether the small amount of missing data (<5%) was randomly distributed. However, in the case where missing data were 24.1% (fear of movement—“Physical activity might harm my back”—MBPNP), we did investigate whether the missing data could bias the results. The missing data were evenly distributed across classes from 22% to 25% (Table 4; class 1: 24.9%, class 2: 23.8%, class 3: 24.2%, and class 4: 22.2%) and considered missing at random and therefore could be ignored.

5. Conclusion

This study adds to the existing literature on the categorization of pain drawings and the association between pain drawings and nonpain symptoms using a data-driven approach. The five- and four-class-model solutions for LBP and MBPNP, respectively, confirmed that more extensive pain is associated with higher activity limitation, more intense pain, and more psychological distress. This study emphasizes that patients referred with site-specific pain often have pain in more than one site of the body and often in a diffuse manner. Furthermore, our study provides a clearer view of pain drawing phenotypes by considering lateralization. Future analyses should investigate to what extent pain drawings can predict and potentially replace questionnaire data on psychological distress. Early identification and differentiation of clinically meaningful subgroups from pain drawings may help inform further examination, clinical management, and prognosis of persistent spinal pain. Future research should explore the potential of pain drawings in clinical contexts, including their prognostic value and diagnostic accuracy, by using these methods.

Conflict of interest statement

The authors have no conflicts of interest to declare.

Appendix A. Supplemental digital content

Supplemental digital content associated with this article can be found online at http://links.lww.com/PAIN/C46.