Brain predicted age difference estimated using DeepBrainNet is significantly associated with pain and function—a multi-institutional and multi-scanner study

1. Introduction

Machine learning has enabled researchers to use high-dimensional magnetic resonance imaging (MRI) datasets to build predictive models of brain aging [15]. Given their high sensitivity, these brain age estimates may capture disease-related brain changes [5,17–19,28,31,34,35], providing proxies to health states. Previously, we and others employed Gaussian Process Regression (GPR) to estimate brain-predicted age differences (brain-PAD) across several chronic pain samples [20,36,37], providing evidence of accelerated brain aging in certain cohorts and pain types [i.e., older adults, knee osteoarthritis (OA), high-impact pain] [37,42]. However, these associations are still debatable, with one investigation showing no difference in brain-PAD between 59 individuals with non-cancer chronic pain and 60 controls [54]. Under the premise that these negative results might owe to the use of an heterogeneous chronic pain sample, Hung et al. (2022) recently compared patients versus controls separately for each chronic pain type and found that brain-PAD significantly differed between OA patients and controls, but not between chronic back pain (CBP) patients and controls [36]. These differences were significant in CBP females but not in CBP males. However, the moderation by sex of these pain-related differences was not explicitly tested. Moreover, their brain age model was trained in a small sample of 812 healthy subjects, and their testing sample was relatively small, with 52 OA and 50 CBP participants. Therefore, replication is needed in larger and more heterogeneous samples.

Here, we implement a convolutional neural network (DeepBrainNet [5]) to investigate the differential association between different chronic musculoskeletal (MSK) pains and brain aging in a much larger (n=660, 169 OA and 170 CBP participants) and more heterogeneous sample from different study sites and ages (i.e., 20–83). Compared to other brain age prediction methods, DeepBrainNet is more suitable for our sample since it was trained on a significantly larger and more heterogeneous dataset (n=11,729) from 18 studies spanning different scanners, ages and locations.

Based on the abovementioned findings where ‘older-appearing’ brains were associated with the presence of pathologies [5,17–19,28,31,34,35], including chronic pain [20,36,37], we hypothesized that DeepBrainNet-based brain-PAD would be significantly greater in either OA and/or CBP participants compared to controls. Under the premise that different neurobiological mechanisms underlie OA and CBP [3], we hypothesize that brain-PAD would also differ among both groups. Moreover, based on the well-known differences in mechanisms of chronic pain among sexes [4,27], specifically in the brain [32], as well as the observed sex differences in brain-PAD [10,22,52], we hypothesized that these differences would be moderated by sex (i.e., women with MSK pain have significantly ‘older-appearing’ brains compared to males or controls).

Finally, although the abovementioned, smaller sampled studies have reported significant associations between brain-PAD and some measures of psychological function [20], experimental pain [20], and clinical pain [36], our larger and more heterogeneous sample furnishes a more comprehensive characterization of the relationship between brain age and the multidimensional experience of pain. We thus explored the association between brain-PAD and an expanded more complete set of clinical pain, sensory or functional variables, and their moderation by sex.

2. Materials and Methods

2.1. Participants and MRI scanners

This is a multicenter study combining seven different musculoskeletal (MSK) pain datasets from eight different MRI scanners (encoded in the variable scanner).

2.1.1. UF/UAB dataset

This was a subsample of a larger multisite observational study carried out at the University of Florida (UF) and at the University of Alabama-Birmingham (UAB) aimed at examining ethnic/race group differences in individuals (>45 years of age) with or at risk for knee osteoarthritis (kOA). The sample included participants with kOA and demographically matched controls. MRI data was collected at the McKnight Brain Institute at the University of Florida using a 3-Tesla Achieva Phillips (Best, the Netherlands) scanner using a 32-channel radio-frequency coil, and at the University of Alabama, Birmingham, with the same MRI system but using an 8-channel head coil. Note that, even though all experimental procedures were identical, and the MRI scanner was the same, each study used a different MRI coil. Therefore, the variable scanner took the values ‘UF Phillips’ and ‘UAB Phillips’. A high resolution, T1-weighted (T1w) turbo field echo anatomical image was collected with TR = 7.0 ms, TE = 3.2 ms, 176 slices acquired in a sagittal orientation, flip angle = 8 degrees, resolution = 1 mm³. Head movement was minimized via cushions positioned inside the head coil.

The study was approved by the University of Florida and the University of Alabama Institutional Review Boards (IRBs) under the Common Rule, which requires the use of single IRB for U.S. based institutions engaged in cooperative research. All participants provided verbal and written informed consent. The study was carried out in accordance with the Declaration of Helsinki.

2.1.2. UF Only dataset

This study included community-dwelling older adults (>60 years) and younger adults (18–30 years) as part of the Neuromodulatory Examination of Pain and Mobility Across the Lifespan (NEPAL) project at the University of Florida (UF). Presence of MSK pain in this sample was determined after the participants completed a standardized pain history interview regarding the presence of pain across several body regions (i.e., head/face, neck, shoulders, arms, hands, chest, stomach, upper and lower back, leg, knees, and feet) using a validated body manikin [21]. Our final sample included participants that reported kOA only, chronic back pain (CBP) only, and demographically matched pain-free individuals (controls). Part of the MRI data was obtained using the same Phillips scanner at UF described for the UF/UAB dataset (the 32-channel configuration). The rest of the MRI data was acquired with a 3T Siemens MAGNETOM Prisma (AG, Erlangen, Germany) scanner (software version VE11C) at UF’s McKnight Brain Institute, where a T1w 3D MPRAGE anatomical image was collected with PAT mode GRAPPA with Phase-Encoding (PE) acceleration factor = 2, 192 sagittal slices, TI = 900 ms, TR = 2300 ms, TE = 2.96 ms, flip angle = 9 degrees, field-of-view (FOV) = 256 × 256 mm and spatial resolution of 1 × 1 × 1 mm. Therefore, scanner took the values ‘UF Phillips’ and ‘UF Siemens’. The study was approved by the UF Institutional Review Board (IRB) and was carried out in accordance with the Declaration of Helsinki.

2.1.3. OpenPain datasets

We used data from four MRI studies with their dataset available in the OpenPain Project database repository (www.openpain.org). These are datasets of different types of MSK pain. To ease their identification in the website, we named them after their folder names in the repository.

OpenPain SALS dataset (folder name: “subacute_longitudinal_study”)

This dataset included 70 participants with subacute back pain (SBP), 26 CBP participants and 26 controls at baseline [59]. This is a longitudinal study with up to five scanning time-points after baseline (at 3.2, 10.4, 31.3, 57.7 and 154. 2 weeks in average). For the present study, we used the data of the controls and CBP participants of the first time-point. Additionally, we added some more CBP participants as follows. Based on the criteria defined by Vachon-Presseau et al. (2016), we classified the SBP participants into “recovered” if their reported severity of pain [based on a Visual Analogue Scale (VAS)] decreased 80% after 56 weeks from baseline [60]. These participants were discarded from our study. Those who did not recover were classified as “persistent” and were deemed as CBP participants after 56 weeks. Thus, their data (MRI and other variables) acquired during their last available time-point beyond 56 weeks (time-point 4 or 5) was used in our study. MPRAGE T1w images were collected at Northwestern University (NU) with a 3 T Siemens Trio, the standard radio-frequency head coil, and the following parameters: voxel size = 1 × 1 × 1 mm, repetition time = 2,500 ms, echo time = 3.36 ms, flip angle = 9°, in-plane matrix resolution = 256 × 256; 160 slices, field of view = 256 mm [59]. Therefore, the variable scanner took the value ‘NU Trio’.

OpenPain PPT (folder name: “placebo_predict_tetreault”)

This dataset included 56 OA participants and 20 controls [57]. MPRAGE T1w images were also collected at Northwestern University with a 3T Siemens Trio using the MRI protocol of OpenPainSALS [2,57]. Therefore, the variable scanner also took the value ‘NU Trio’. Note that fifty two (52) out of the 56 OA participants in this study were previously used by Hung et al. (2022) to explore differential changes in brain-PAD associated with chronic pain [36].

OpenPain CBPR dataset (folder name: “cbp_resting”)

This dataset included 34 CBP participants and 34 matched controls. The MRI protocol for this dataset is unavailable. We assumed they were acquired at Northwestern University using the same protocols of the SALS and PPT dataset. However, since their brain images are only available in their skull-stripped version, we set the value of scanner to ‘NU Trio SS’. The exceptions were the participants with IDs ‘healthy16’ and ‘healthy17’ that had whole brain MRIs and thus were assigned to scanner = ‘NU trio’.

The OpenPain ACPS (folder name: “AccumbensChronicPainSignature”)

This is a two-time-point study (median follow-up of 59.5 weeks), including 29 CBP participants 16 persistent SBP, 19 recovery SBP, and 33 controls [43]. Like with OpenPain SALS, recovered SBP participants were discarded and persistent SBP participants were deemed CBP and their 56 weeks+ follow-up data used. MPRAGE T1w images were collected at Yale University using a Siemens 3-T Trio B magnet with a 32-channel head coil and TR = 1,900 ms, TE = 2.52 ms, flip angle = 9°, and matrix 256 × 256 with 176 slices (1 mm thick). For this dataset, images available in the OpenPain repository were already skull-stripped [43]. Thus, scanner = ‘YU Trio’.

The OpenPain BNCM (folder name: “BrainNetworkChange_Mano”)

This dataset included 41 CBP participants and 56 controls. Images were acquired using a 3-T MRI Scanner (3T Magnetom Trio with TIM system; Siemens, Erlangen, Germany) with a standard 12-channel phased array head coil either at Addenbrooke’s hospital (Cambridge, UK) or CiNet (Osaka, Japan). A high-resolution three-dimensional T1w image was collected using a MPRAGE pulse sequence. For the participants in the UK, TR = 2300 ms, TE = 2.98 ms, time of inversion = 900 ms, FA = 9 degrees, BW = 240 Hz, FOV = 256 × 256 mm, 176 sagittal slices of 1mm slice thickness with no inter-slice gap, acquisition matrix = 256 × 256. For the participants in Japan, TR = 2250 ms, TE = 3.06 ms, time of inversion = 900 ms, FA = 9 degrees, BW = 230 Hz, FOV = 256 × 256 mm, 208 sagittal slices of 1mm slice thickness with no inter-slice gap, acquisition matrix = 256 × 256 [44]. Therefore, scanner took the values ‘Addenbrooke Trio’ for 17 CBP patients and 17 controls; and ‘CiNet Trio’ for CBP 24 patients and 39 controls. As reported in [36], the CBP participants in this study were also used by Hung et al. (2022) to explore differential changes in brain age difference associated with chronic pain [36].

2.3. DeepBrainNet-based brain age prediction

Developed by Bashyam et al. (2020) [5], DeepBrainNet is a convolutional neural network-based brain age prediction method. It is built based on the inception-resnet-v2 framework [56] and uses a 2D convolutional architecture. DeepBrainNet was trained using T1w MRI images from 11,729 individuals (ages 3–95 years) from a diverse range of geographic locations, scanners, acquisition protocols, and studies, and tested in an independent sample of 2,739 individuals. Features for the DeepBrainNet are calculated as follows. First, the T1w scan needs to be skull-stripped (i.e. extracranial tissues must be removed using image pre-processing methods, so that only gray and white matter, as well as CSF, are kept). Second, the skull-stripped image has to be spatially normalized to the 1-mm isotropic voxel FSL skull-stripped T1w template using a 12-parameter linear affine transformation. For training, each of the skull-stripped MRI image was divided into 80 2D slices (centered on the z = 0 plane in MNI coordinates) and considered as an independent sample, resulting in a training set of 1 million images. To obtain a final age prediction for a test sample, each of 80 slices of the test scan is input to the trained model independently and the median prediction is calculated as the subject’s predicted brain age. To obtain skull-stripped images in our sample, we used smriprep¹, the portion that process the anatomical T1w images in fmriprep [26]. Briefly, the T1w image was corrected for intensity non-uniformity using N4BiasFieldCorrection [58] distributed with ANTs 2.2.0 (Avants et al. 2008 [1], RRID:SCR_004757), and skull-stripped with a Nipype implementation of the antsBrainExtraction.sh workflow from ANTs, using OASIS30ANTs as target template. The skull-stripping step was omitted for those images already available in their skull-stripped version (OpenPain CBR and OpenPain ACPS).

2.4. Measures characterizing experimental pain, function and clinical pain

For each brain age prediction method, we explored the association between brain-PAD and several variables characterizing clinical pain, experimental pain (i.e., Quantitative Sensory Testing [QST]) and function (psychosocial, physical and cognitive). These variables largely came from the UF and UAB participants, but some were also available from the OpenPain datasets. Table 1 shows for which datasets each of these variables were measured.

Table 1.

Variables characterizing clinical pain, experimental pain and function in the datasets.

	UF/UAB	UF	CBPR	SALS	ACPS	BNCM
Experimental Pain/QST
Heat Pain (P.) TS Index	✓
Heat P. Sensitivity Index	✓	✓
Cold P. Rating Index	✓	✓
Punctate P. Sensitivity Index	✓	✓
Punctate P. TS Index	✓	✓
Pressure P. Index	✓	✓
CPM-During	✓
CPM-Post	✓
Psychosocial Function
CSQ-R-Catastrophizing	✓	✓
CSQ-R-Coping	✓	✓
CSQ-R-Active Coping	✓	✓
CSQ-R-Passive Coping	✓	✓
CSQ-R-Distancing	✓	✓
CSQ-R-Ignoring	✓	✓
CSQ-R-Prayer	✓	✓
CSQ-R-Distraction	✓	✓
IVC-Active Coping	✓
IVC-Passive Coping	✓
PANAS-Negative Affect	✓	✓		✓
PANAS-Positive Affect	✓	✓		✓
MSPSS	✓
Somatization (PHQ-15)	✓
PROMIS-Anxiety	✓
PROMIS-Depression	✓
PROMIS-Sleep	✓
PSQI Total		✓
PSQI Duration	✓
Severity of Insomnia	✓
BDI			✓	✓	✓	✓
Physical and Cognitive Function
SPPB Total Score	✓	✓
MoCA Total Score	✓	✓
Clinical Pain
Number of Pain Sites	✓	✓
GCPS-Pain Intensity	✓	✓
WOMAC-Pain	✓	✓
PD-Q Total	✓	✓		✓
SF-MPQ-2-Continuous	✓	✓
SF-MPQ-2-Intermittent	✓	✓
SF-MPQ-2-Neuropathic	✓	✓
SF-MPQ-2-Affective	✓	✓		✓
SF-MPQ-2-Total	✓	✓
KL Index	✓
Pain Length	✓	✓	✓	✓	✓
Pain Duration		✓	✓	✓	✓	✓

Note. None of these variables were measured in OpenPain PPT. TS: Temporal Summation. CSQ-R: Coping Strategies Questionnaire-Revised. IVC: In-Vivo Coping. PANAS: Positive and Negative Affect Scale. MSPSS: Multidimensional Scale of Perceived Social Support. BDI: Beck Depression Inventory. PHQ-15: Patient Health Questionnaire Somatic Symptom Severity Scale Item 15. PROMIS: Patient-Reported Outcomes Measurement Information System. PSQI: Pittsburgh Sleep Quality Index. SPPB: Short Physical Performance Battery. MoCA: Montreal Cognitive Assessment. SF-MPQ-2: Short Form McGill Pain Questionnaire-Revised. GCPS: Graded Chronic Pain Scale. PD-Q: Pain Detect Questionnaire. WOMAC: Western Ontario and McMaster Universities Osteoarthritis Index – Pain. KL: Kellgren-Lawrence.

2.4.1. Clinical pain variables

Number of Pain Locations

UF/UAB and UF participants were asked to indicate areas where they experienced pain, including head, neck, shoulders, chest, stomach, upper back, lower back, arms, hands, knees, legs and feet. The total number of pain locations was used for the analysis.

For the participants in OpenPain SALS, we used the radiculopathy scores that were quantified as the total of pain locations that patients had shaded in with pencil on the Short Form McGill Pain Questionnaire (SF-MPQ) form [45]. More details about how this questionnaire was administered in OpenPain SALS can be found elsewhere [2,14].

Graded Chronic Pain Scale (GCPS).

The GCPS is a 7-item scale that measures characteristic pain intensity and pain interference over the past 6 months. Participants are asked to rate their current, average, and worst pain on a 0 “no pain” to 10 “worst pain imaginable” numeric rating scale (NRS). Ratings were averaged and multiplied by 10 to calculate a characteristic pain intensity score (range: 0–100), with higher scores indicating greater pain intensity [39]. This was assessed in UF/UAB and UF participants.

Western Ontario and McMaster Universities Osteoarthritis Index - Pain (WOMAC-pain)

The WOMAC-pain subscale assesses pain experienced in the lower limbs during various activities [7]. Items are rated on a 5-point scale, with higher scores indicating greater levels of pain during activities with scores ranging from 0–20. This was assessed in UAB/UF and UF studies.

The Pain Detect Questionnaire (PD-Q)

The PD-Q [30] is a reliable screening tool with high sensitivity, specificity and positive predictive accuracy to assess the likelihood of a neuropathic pain component in patients. Scores on the PD-Q range from 0 to 38, with a score of 12 and higher generally considered as neuropathic pain. This questionnaire was administered in the UAB/UF, UF and OpenPain SALS studies.

Short Form McGill Pain Questionnaire-Revised (SF-MPQ-2)

The SF-MPQ-2 is used to measure the quality and the intensity of pain [24]. It comprises Continuous pain, Intermittent pain, Neuropathic-type pain, and Affective experiences subscales. Each of 22 pain descriptors is rated on a 0 “no pain” to 10 “worst pain ever” scale within the past week, and a Total sum score is calculated for each subscale. This questionnaire was administered in the UF and UF/UAB studies, and the Continuous and Affective subscales were also available in the OpenPain SALS database.

Kellgren-Lawrence Index

Radiographs were obtained for participants in the UF/UAB study only to determine degree of joint pathology according to the Kellgren-Lawrence (KL) criteria [38]. Grades range from 0 to 4 with higher grades indicating worse joint pathology.

Pain Length and Pain Duration

UF/UAB participants were asked to report how long they had been experiencing knee pain (i.e., Pain Length): 1) Less than 6 months, 2) 6 months to 1 year, 3) 1 to 3 years, 4) 3 to 5 years, 5) Greater than 5 years. Participants from the UF, OpenPain CBPR, SALS and ACPS datasets self-reported the number of years that they had experienced back pain (i.e., Pain Duration). Thus, for these studies, years were also classified according to the Pain Length categories as in the UF/UAB dataset.

Presence of MSK chronic pain

UAB/UF and UF only participants that self-reported pain for the past 3 months on most days with a GCPS pain intensity of 40 or higher in a 0–100 scale were deemed as chronic pain participants. For the OpenPain datasets, classification of chronic pain participants is provided in the metadata files found alongside the MRI data in the dataset’s repository. The criteria used to classify a participant as having chronic pain in each OpenPain dataset is described in its corresponding paper(s) [43,44,57,60], including those subacute back pain participants at baseline who did not recover after 56 weeks (“persistent”) who we classified as CBP participants. In our analyses, we created the variable MSK pain presence taking the values ‘yes, or ‘no’.

2.4.2. Experimental pain/quantitative sensory testing (QST) variables

The UF only and UF/UAB cohorts completed a multimodal QST session approximately one week prior to MRI data collection. Procedures were standardized and completed by trained research staff as follows.

Thermal pain

Heat Temporal Summation (TS) Index

A contact heat-evoked potential stimulator thermode (Medoc Pathway; Ramat Yishai, Israel) was used to deliver five sequential heat pulses in three separate trials (i.e., 44°C, 46°C, and 48°C). Each trial began at a baseline temperature (35°C) and rapidly increased (20°C/sec) to the target temperature. Participants were asked to rate their pain at the peak of each heat pulse on a 0 “no pain” to 100 “most intense pain imaginable” Numerical Rate Scale (NRS). The trial was ended if the participant provided a pain rating of 100 or requested to stop. Temporal summation (TS) was calculated as: maximum pain rating – first pain rating within each trial. TS values were standardized and averaged across body sites and temperatures to create a Heat Temporal Summation Index, as in our previous work [12]. Higher values indicate greater heat TS.

Heat Pain Sensitivity Index

UF/UAB dataset:

Heat stimuli were applied to the medial joint line of the most painful knee and on the ipsilateral ventral forearm using a 16×16 thermode (Medoc Pathway Thermal Sensory Analyzer; Ramat Yishai, Israel). Heat pain threshold was considered the first sensation of pain and heat pain tolerance was the point at which pain could no longer be tolerated. Each trial began at a baseline temperature of 32°C and gradually increased (0.5°C/sec) until ended by the participant pressing a button to stop the trial. Pain was rated following each trial of heat pain threshold and tolerance on a 0 “no pain” to 100 “most intense pain imaginable” NRS. The mean of three trials was used for analysis. Heat pain threshold, heat pain tolerance, and pain ratings from all heat TS pulse series (described above) at both body sites were standardized and averaged to compute a Heat Pain Sensitivity Index, with higher values indicating greater heat pain sensitivity.

UF only dataset:

Heat stimuli were applied to the thenar eminence and first metatarsal using a 30×30 thermode (Medoc Pathway Thermal Sensory Analyzer; Ramat Yishai, Israel), with a starting temperature of 32°C and gradually increasing at a rate of 1°C/sec until the participant reported the stimulus as first painful (heat pain threshold). Participants rated the pain intensity of each trial on a 0–100 NRS. Three trials were conducted at each body site, and the average obtained for each test. Heat pain threshold and pain ratings were standardized and averaged to compute a Heat Pain Sensitivity Index, with higher values indicating greater heat pain sensitivity.

Cold Pain Rating Index

UF/UAB:

During a conditioned pain modulation trial, participants were asked to immerse their open hand, up to the wrist, into a 12°C water bath (Neslab, Portsmouth, NH, USA). Pain was assessed at 30 seconds on a 0–100 NRS. The procedure was repeated two times, separated by a 10-minute recovery period incorporating a heating pack for one minute. Pain ratings for both trials were standardized and averaged for an overall Cold Pain Rating Index, with higher values indicating greater cold pain sensitivity.

UF only:

Cold stimuli was delivered to the thenar eminence and the first metatarsal using a 30×30 thermode (Medoc Pathway Thermal Sensory Analyzer; Ramat Yishai, Israel), with temperature starting at 32°C and decreasing at a rate of 1°C/sec with a cutoff value of 0°C. Participants were asked to indicate when the stimulus first became painful and rate the pain on a 0–100 NRS, for three trials at each body site. An average cold pain rating was computed across the three trials. Cold pain ratings were then standardized and averaged across testing sites for a Cold Pain Rating Index, with higher values indicating greater cold pain sensitivity.

Mechanical pain.

Punctate Mechanical Pain

Mechanical punctate stimuli were applied using a nylon monofilament (Touchtest Sensory Evaluator 6.65) calibrated to bend at 300g of pressure. Stimuli were applied to the patella of the most painful knee and the dorsal aspect of the ipsilateral hand (UF/UAB) and to the thenar eminence and first metatarsal (UF). Pain was assessed on a 0–100 NRS after a single contact, and then following a series of 10 contacts, delivered at a rate of 1/sec. Trials were repeated two times at each body site. Pain ratings from the single contacts from each body site and trial were standardized and averaged to calculate a Punctate Pain Sensitivity Index, with higher values indicating greater sensitivity.

Punctate Pain TS Index

To calculate this index, pain ratings from the single contact were subtracted from pain ratings following the series of 10 contacts. Difference values were standardized and averaged across body sites and trials, with higher values indicating greater temporal summation.

Pressure Pain Index

Pressure pain threshold was assessed at the medial and lateral joint lines of the most painful knee, the ipsilateral quadriceps, and trapezius muscle, with site order randomized, in the UF/UAB sample, and at the trapezius and the quadriceps in the UF sample using a handheld digital pressure algometer (AlgoMed, Medoc, Ramat Yishai, Israel). Participants were asked to press a button when the sensation first became painful. To maintain participant safety, a limit of 600 kPa (knee sites) and 1000 kPa (quadriceps and trapezius) was imposed. Three pressure pain threshold values were averaged for each body site. These values were standardized and combined to calculate a Pressure Pain Index for each sample. Values were reversed prior to combining so that greater scores represent more pain sensitivity.

Conditioned Pain Modulation (CPM).

CPM was assessed using pressure pain applied to the left trapezius and cold-water immersion (as described above). Pressure pain threshold (PPT) was assessed at the trapezius, then the participant put their hand into the cold-water bath, submerged up to the wrist. At 30 seconds, cold pain was assessed followed by PPT. At one minute the participant removed their hand from the cold-water bath and PPT was assessed immediately. After the first trial, a warm pack was placed on the participant’s hand for one minute, and the trial was repeated following a 10-minute rest period. The difference between PPT pre-immersion and PPT during-immersion was averaged across both trials to calculate CPM-During. The difference between PPT pre-immersion and PPT post-immersion was averaged across both trials to calculate CPM-Post. Positive values indicate a CPM response.

2.4.3. Psychosocial function variables

Coping Strategies Questionnaire-Revised (CSQ-R)

UF only and UF/UAB cohorts completed the CSQ-R [50,51] assessed typical coping responses to pain using 27 items divided across 6 types of coping responses: 1) Catastrophizing, 2) Coping self-statements, 3) Distancing, 4) Ignoring, 5) Prayer, and 6) Distraction. Items are rated on a 7-point Likert-type scale from 0 “never do that” to 6 “always do that”, with higher scores indicating greater use of that strategy. Active coping is computed as the mean scores of the Coping self-statements, Distancing, Ignoring, and Distraction subscales. Passive coping is computed as the mean scores of the Catastrophizing and Prayer subscales.

In-Vivo Coping (IVC)

UF/UAB participants completed the IVC [25] to assess situational pain coping strategies following a quantitative sensory testing (QST) battery, including passive (e.g., pain catastrophizing; “I felt that if the pain got any worse, I wouldn’t be able to tolerate it”), and active (distraction; “I thought of other things to get my mind off of the pain”) coping statements. Items were rated on a 1 “not at all” to 5 “very much” point scales. Active and passive coping subscale scores are the summed average of the items within each domain, with higher scores indicating greater use of that type of coping strategy.

Positive and Negative Affect Scale (PANAS)

The PANAS is comprised of 20-items, 10 positive-valence (i.e., interested, excited, strong, enthusiastic, proud, alert, inspired, determined, attentive, active) and 10 negative-valence (i.e., distressed, upset, nervous, scared, hostile, irritable, ashamed, jittery, afraid, guilty). Items are rated on a 5-point scale ranging from 1 “very slightly or not at all” to 5 “extremely” [62]. Higher scores on positive items indicate higher positive affect (PA), while higher scores on negative items indicate higher negative affect (NA). PANAS was assessed in the UF/UAB, UF and OpenPain SALS studies.

Multidimensional Scale of Perceived Social Support (MSPSS)

The MSPSS [65] is a brief measure of subjective social support rated on a 7-point scaled asking individuals to rate the perceived adequacy of support they receive from family, friends, and significant other. Higher scores indicated greater perceived social support. This was assessed in the UF/UAB study only.

Patient Health Questionnaire Somatic Symptom Severity Scale Item 15 (PHQ-15)

The PHQ-15 [41] assessed the degree to which participants are currently distressed about 15 common somatic symptoms. Higher scores indicating greater somatic sensitivity. Somatization was assessed in the UF/UAB study only.

Patient-Reported Outcomes Measurement Information System (PROMIS)

The Depression Short Form (PROMIS-D-SF) [13] consists of 8 items to assess depressive symptomology, with higher scores indicating more depressive symptoms. The PROMIS Anxiety Short Form (PROMIS-A-SF) consists of 7 items rated on a 5-ponit scale. Higher scores indicate greater anxiety type symptoms. These measures were administered in the UF/UAB study only.

Sleep Symptoms

UF study participants completed the Pittsburgh Sleep Quality Index (PSQI) [11] scale which assessed sleep quality over a 1-month time interval. The instrument is used to measure the quality and patterns of sleep in seven domains: 1) subjective sleep quality, 2) sleep latency (i.e., the time it takes to fall asleep), 3) sleep duration (PSQI-Duration in Table 1, encoded as >7hrs=0, 6–7hrs=1, 5–6hrs=2, <5hrs=3), 4) habitual sleep efficiency (the ratio of total sleep time to time in bed), 5) sleep disturbances, 6) the use of sleep-promoting medication (i.e., prescribed or over- the-counter), and 7) daytime dysfunction over the last month. Each of the seven domains is a rated on a 0 to 3 scale. The sum of the components produces a global score ranging from 0 to 21, where a higher score indicates worse sleep quality [11].

UF/UAB study participants completed several questionnaires to assess sleep as follows. 1) The Severity of Insomnia questionnaire assessed difficulty falling asleep, staying asleep and problems waking up too early, and satisfaction with sleep pattern [0 (none/very satisfied) to 4 (very severe/very dissatisfied)], as well as others’ perception of how sleep-related impairments impacted their quality of life [0 (not noticeable) to 4 (very much noticeable)], worriedness/distress about sleep problems [0 (none) to 4 (very much)] and sleep-related interference with daily functioning [0 (none) to 4 (very much)]. Higher scores (range: 0–28) indicated greater insomnia symptoms. 2) The PSQI-Duration subscale. 3) The PROMIS Sleep-Related Impairment scale [49,64] consisted of 8 items which assessed self-reported perceptions of alertness, sleepiness, and tiredness during usual waking hours, and the perceived functional impairments during wakefulness associated with sleep problems and impaired alertness over the past 7 days. Higher scores indicate greater sleep impairment.

Beck Depression Inventory

The Beck Depression Inventory (BDI) is a 21-item, self-report rating inventory that measures characteristic attitudes and symptoms of depression [6]. This inventory was available for the OpenPain CBR, SALS and ACPS databases.

2.4.4. Physical and cognitive function variables

Short Physical Performance Battery (SPPB).

The SPPB [33] consists of three measures of lower-extremity function: standing balance (side-by-side, semi-tandem, and tandem stance), 4-meter walking speed, and ability to rise from a chair. Each task is rated on a 0–4 scale, with increasing scores indicating better physical performance. Total scores range from 0–12. This was assessed in UF/UAB study and UF participants.

Montreal Cognitive Assessment (MoCA)

The MoCA [40] was administered to assess global cognitive abilities including short-term memory, orientation, executive function, language abilities, animal naming, abstraction, attention, clock- drawing test. Scores on the MoCA range from 0 to 30, with a score of 26 and higher generally considered normal global cognition. [40]. This was assessed in UF/UAB and UF study participants.

2.5. Statistical Analysis

Missing data was treated using pairwise elimination. Statistical significance was set to α = 0.05 after correcting p-values using False Discovery Rate (FDR) [8]. To ensure normality, for every model (which entailed a specific subset of the whole dataset), we applied a rank-based inverse normal transformation to the dependent variable (brain-PAD) using the ‘Blom’ method with parameter c= 3/8 [23]. After fitting the models, we applied the Shapiro-Wilk test of composite normality (with unspecified mean and variance) on the residuals (for Platykurtic distributions; while the Shapiro-Francia test was used for Leptokurtic distributions) to test whether the normality assumption required for linear models was fulfilled [53].

Given the multisite and multi-scanner nature of our sample, previous to testing our proposed hypotheses, we considered informative to report the extent to which this heterogeneity could affect the estimation of the brain-PAD. Moreover, we wanted to evaluate whether the pain-related differences were consistent across scanners, i.e., if there was not significant interaction between MSK pain presence and scanner. To that end we tested whether there was a significant effect of scanner on MSK pain-controls differences in brain-PAD, we fitted the linear model, in Wilkinson’s notation, brain-PAD ~ MSK pain presence * scanner + sex + age, where sex took the values ‘male’ and ‘female’, and age is the chronological age. This analysis would also inform about the need to add terms accounting for pain-by-scanner interactions in the subsequent analyses.

We hypothesized that brain-PAD would be significantly different among OA and CBP participants and that it would be greater in OA and/or CBP participants compared to controls (H_type). We also hypothesized that these differences would be moderated by sex (H_type-by-sex). To test these hypotheses, we fitted a linear mixed model brain-PAD ~ pain type * sex + age + (1 | scanner); where pain type was a categorical variable taking the values ‘control’, ‘oa’ and ‘cbp’ (i.e., an ANCOVA with random effects). Note that race was not included as a covariate in any of these models since it was not available for the OpenPain ACPS, OpenPain CBPR and OpenPain PPT databases. After fitting this model, we were first interested in evaluating whether the interaction effect pain type : sex was significant (H_type-by-sex). In case of a significant interaction, we were then interested in evaluating a set of contrasts of interest to understand the direction of the effects, i.e., simple effects of pain type at the levels of sex (H_type, by sex) and simple effects of sex at the level of pain type. Also using the coefficients of this model, we were also interested in replicating the above-mentioned reports by Hung et al., (2002) of a differential effect of pain among MSK pain types [36], i.e., the effects of pain type after averaging across the levels of sex (H_type, but marginalizing sex). In the case of no significant interaction, we fitted the reduced model without interaction brain-PAD ~ pain type + sex + age + (1 | scanner) instead, and evaluated the effect of pain type (H_type).

We were also interested in replicating the difference in brain-PAD between MSK pain (irrespective of type, namely H_MSK) and controls we previously observed in a smaller sample [20]. Leveraging the abovementioned linear mixed model, we tested this by evaluating the statistical significance of the average of the marginal means of the brain-PAD across OA and CBP minus the marginal mean of the brain-PAD of the controls.

We were also interested in exploring the association between brain-PAD and a set of clinical pain, sensory or functional variables (H_var), and whether they would be moderated by sex (H_var-by-sex)². Similar to the testing of H_type and H H_type-by-sex, a general framework to explore these associations is to fit the linear mixed model brain-PAD ~ variable of interest * sex + age + race + (variable of interest | scanner), where variable of interest is each of the clinical pain or function variable. Note that we added the categorical variable race, taking the values ‘African American’, ‘Caucasian’, ‘Hispanic’, ‘Asian’, and ‘Pacific Islander’, since the variables of interest were only present in samples for which race information was available. The only exceptions were for Pain Duration and the Beck Depression Inventory, because they were available in OpenPain ACPS and OpenPain CBPR, which had no race information. Also note that we modeled a random intercept and random effects of variable of interest per study site (scanner can be used to encode study site), the latter to account for possible differences in the way these variables were acquired at different studies. We first tested the significance of the interaction term (variable of interest : sex), i.e., H_var-by-sex, and then, in case of significance, we tested the simple effects of variable of interest at level of sex (H_var). In case the interaction term was not significant, we tested the effect of variable of interest (H_var) in a simple model not including the interaction term. For the experimental pain and function variables, the regressions were performed using the whole sample, while for the clinical pain, the regressions were performed only those participants having MSK pain. Note that, while a moderation-by-sex analysis allows to compare the slopes of the associations across sexes, a comparison of the correlation would allow to compare how different is the strength of the association between the variable of interest and brain-PAD. Thus, we also report the partial correlations between each variable of interest and brain-PAD for each sex and their comparison using the z-test for the comparison between correlations.

Effects sizes were reported using the Cohen’s f², which measures the relative variance explained by the effect when added to the regression model (0.1² ≤ f² < 0.25² for small effects, 0.25² ≤ f² < 0.4² for medium effects and f² > 0.4² for large effects. For pairwise comparisons, we additionally reported the difference in marginal means, namely ΔPAD, and its Standard Error (SE).

Mixed models were fitted by maximizing their Likelihood using the ‘Quasi-Newton’ optimizer, tolerance of 1e-16, step size tolerance of 1e-12 and maximum 10,000 iterations. All analyses were re-run after removing those measurements deemed outliers, based on their Cook’s distance being 3 times higher than their sample average [47], and their results are presented in the Supplemental Materials.

3. Results

3.1. Final sample size and participants demographics

All preprocessed MRI images were submitted to a careful quality control procedure. All raw images, segmented, brain masked, and normalized images were visually inspected. Those having poor signal-to-noise (SNR), inaccurate brain extraction or poor spatial normalization were discarded from the study (see Table S1 in the Supplemental Materials for details). This led to a final sample of 321 controls and 339 participants with MSK pain, of which 169 had OA and 170 had CBP, for a total of 660 individuals across three groups. Detailed demographic information is shown in Table 2 and 3. There were no significant differences in sex distribution by group (χ² test = 2.37, p = 0.31), but there was a significant difference in sex by scanner (χ² = 22.3, p = 0.0022). There were also no significant age differences between MSK pain presence and controls. However, because OA predominantly manifests in middle-aged and older adults, the minimum age of the OA participants was 45 years; while it was 19 and 18 years for the controls and CBP participants (p < 1e-20, Welch’s ANOVA effect of pain type on age). Also, females were slightly older than males in our sample (mean ages 51.6 and 48.1 years, respectively, p = 0.0045) due males having an age distribution slightly more negatively skewed than that of females (i.e., fewer females in the 20–40 year range and fewer males in the 50–70 year range).

Table 2.

Distribution of participants by pain type, sex, race and study sites/scanners.

		Pain Type				Sex		Race
		Control	OA	CBP	MSK (OA+CBP)	M	F	AA	C	H/O
Scanner	NU Trio	42	54	48	102	70	74	28*	28*	12*
	NU Trio SS	32	0	30	30	33	29	n/a	n/a	n/a
	Yale Trio	33	0	40	40	37	36	n/a	n/a	n/a
	UF Achieva	121	68	11	79	69	127	54	131	11
	UAB Achieva	32	47	0	47	28	51	43	36	0
	UF Prisma	13	0	2	2	8	5	1	12	3
	Adden. Trio	37	0	22	22	35	24	n/a	n/a	n/a
	CiNET Trio	17	0	17	17	11	23	n/a	n/a	n/a
	TOTAL	321	169	170	339	291	369	126	204	26
Sex	Male	148	66	77	143			50	77	13
	Female	173	103	93	196			76	127	13
Race	AA	32	70	24	94	76	50
	C	132	45	27	72	115	89
	H/O	16	0	10	10	13	13

Note. Adden: Addenbrook. OP: OpenPain. OA: Osteoarthritis Pain. CBP: Chronic Back Pain. MSK: Musculoskeletal pain.

^*Race information is not available for OpenPain CBPR, PPT and BNCM datasets. n/a: Not available. M: male. F: Female AA: African American. C: Caucasian. H/O: Hispanic or Other. Race category “Other” included 3 Asian and 3 Pacific Islanders.

Table 3.

Distribution of age across pain types and sex.

Statistics	Pain Type				Sex
	Control	OA	CBP	MSK (OA+CBP)	Male	Female
Mean age	48.9	57.4	44.8	51.1	48.1	51.6
SD age	17.3	7.5	13.7	12.7	16.0	14.3
Minimum age	19.0	44.0	18.0	18.0	18.0	18.0
Maximum age	82.9	83.0	85.3	85.3	85.3	83.0
Effect of grouping variable p-value	pain type p < 1e-20			MSK pain presence p = 0.063	sex p = 0.0045

Note. OA: Osteoarthritis Pain. CBP: Chronic Back Pain. MSK: Musculoskeletal pain. Due to heteroscedasticity, Welch’s ANOVA was used to estimate the effect of pain type. For the rest of the comparisons, both Welch’s ANOVA and ANOVA led to almost identical results.

The average severity of clinical pain (scaled from 1–100) in the total MSK pain sample, as reported using a Visual-Analogue-Scale (VAS) for all studies except the UF study that used a Numerical Rating Scale (NRS) and the OpenPain PPT that did not have any available, was 45.7, and distributed according to a minimum, 25 percentile, median, 75 percentile and maximum of 0, 44, 58.2, 73.3 and 100, respectively. Table 4 summarizes other pain characteristics of the sample.

Table 4.

Pain characteristics in the total sample

Pain characteristic	Mean ± SD	Min-Max	Studies where measured
Clinical pain severity	57.6 ± 21.6	0–100	All studies except OP PPT
Number of Pain Locations	4.9 ± 3.5	1–21	UF/UAB, UF and OP SALS
GCPS pain intensity	62.95 ± 16.4	40–100	UF/UAB and UF
WOMAC pain	8.0 ± 4.3	1–20	UF/UAB and UF
Pain Duration (years)	9.5 ± 9.6	<1 – 56	UF/UAB and UF

Note. OP: OpenPain

3.2. Brain age predictions

In the prediction of brain age, the DeepBrainNet model yielded a mean absolute error (MAE_controls) = 6.69, 95%CI [6.16, 7.25] years for the controls (n = 321) and MAE_MSK = 6.19, 95%CI [5.68, 6.75] years for the MSK pain participants (n = 339). Also, the correlations between the chronological and predicted brain ages were high and highly significant (i.e., r = 0.88 with p < 1e-20 for controls and r = 0.79 with p < 1e-20 for MSK pain participants) (See Figure 1). For the complete dataset (n = 660), MAE_total = 6.43, 95%CI [6.07, 6.82], and r = 0.86 with p < 1e-20. An independent sample t-test revealed that the difference MAE_MSK - MAE_controls = −0.5 years was not significantly different from zero, with a 95% CI (obtained with 10,000 bootstraps) of [−0.26, 1.26] years.

Figure 1.

Brain age predictions for controls and MSK pain participants. MAE: Mean Absolute Error. Circles, diamonds and squares correspond to controls, OA and CBP participants, respectively.

3.3. Brain-predicted age difference by scanner

We explored the effect of this important confounder by fitting the linear model brain-PAD ~ MSK pain presence * scanner + sex + age. We found that the MSK pain presence : scanner interaction term was not significant (p = 0.69). On the other hand, the effect of scanner was very significant (p = 3.55e-20, Cohen’s f² = 0.18 > 0.4²), while the effect of MSK pain presence was also significant [p = 0.00073, 0.1² ≤ Cohen’s f² = 0.017 < 0.25², with difference in the marginal means of brain-PAD (ΔPAD) between MSK individuals and controls of 1.7 years and Standard Error (SE) of 0.5 years], revealing that it would be impossible to detect any difference in brain-PAD between MSK individuals and controls without removing the confounding effect of scanner. These can be appreciated in Figure 2.

Figure 2.

Distribution of the predicted age difference (brain-PAD) predicted by DeepBrainNet distributed across MSK pain presence and scanner. In the top panel, the significant large main effect of scanner (p = 5.10e-21 with Cohen’s f² = 0.19 > 0.4²; modeled in brain-PAD ~ MSK pain presence * scanner + sex + age) is visually appreciated in the great variability in the brain-PAD among scanners. The bottom panel depicts the distribution of brain-PAD values after correcting for the main effect of scanner, sex and age, allowing to depict a significant (but small) effect of MSK pain presence [p = 0.0018, 0.1² ≤ Cohen’s f² = 0.015 < 0.25², and ΔPAD (SE) = 3.1 (0.2) years] that was recovered after the correction. There was no significant MSK pain presence by scanner interaction in this model (p = 0.69 for the MSK pain presence:scanner interaction effect).

3.4. Pain-related differences in brain-predicted age

The interaction term pain type : sex (corresponding to the hypothesis H_type-by-sex) was not statistically significant (p > 0.05) in the model brain-PAD ~ pain type * sex + scanner + age + (1|scanner). We thus fitted the model without an interaction term (H_type) and found a significant main effect of pain type (p = 1.0e-5) with a small effect size (Cohen’s f² = 0.04 ≤ 0.25²). Post-hoc pairwise differences revealed that brain-PAD of OA participants was significantly higher than that of CBP participants [p = 0.0022 with ΔPAD (SE) = 2.7 (0.79) years] and controls [p = 5.5e-6 with ΔPAD (SE) = 3.1 (0.64) years]. There was no significant difference between CBP participants and controls. Finally, using the contrast comparing OA+CBP and controls, we found that brain-PAD was significantly higher in the whole group of MSK pain participants compared to the controls [H_MSK; p = 0.00048 with ΔPAD (SE) = 1.75 (0.5)]. This is all summarized in Figure 3. The figure also shows that, after discarding outliers, the effect size of pain type became large (Cohen’s f² = 0.1 ≥ 0.25²), with stronger evidence for all effects [e.g., OA vs. CBP: p = 6.5e-7 with ΔPAD (SE) = 3.56 (0.68) years, OA vs. control: p = 9.0e-11 with ΔPAD (SE) = 3.69 (0.55) years]. For all models, rejection of composite normality failed after correcting for multiple comparisons (p > 0.05).

Figure 3.

Differential effect of sex across pain types and main effect of pain type on brain-PAD. Panel A) Plotted brain-PAD values were adjusted for effects of all independent variables in the mixed model brain-PAD ~ pain type * scanner + sex + age + (1|scanner) except pain type and sex. The figure depicts the Bonferroni corrected p-values of all contrasts of interest, i.e., the simple effects of sex at the levels of pain type, the simple effects of pain type at the levels of sex, as well as the p-values of the differences in brain-PAD, averaged across sexes, between the levels of pain types, and pain versus no pain (i.e., the marginal main effect of MSK pain presence). Panel B) same results after removing outliers. Within the violin plots, the shaded are is the interquartile region, the white dot indicates the median and the black horizontal line is the mean. DoF: Degrees of Freedom. ΔPAD (in years): difference in brain-PAD across groups. SE (in years): Standard Error.

With the goal of replicating the reports by Hung et al. (2022), we also evaluated the differences between each MSK type and controls for each sex subsample (i.e., H_type for each sex). For a more complete analysis, we also included the OA versus CBP comparison for each sex. After Bonferroni correction across pairs, we found a significant difference between the OA and control groups for females [p = 0.0046 ΔPAD (SE) = 2.62 (0.82) years] and for males [p = 0.00061 ΔPAD (SE) = 3.79 (1.01) years], as well as a significant difference between the OA and CBP groups for males [p = 0.0049 ΔPAD (SE) = 3.77 (1.19) years], but not for females. When removing outliers, we found larges and more significant differences for all four comparisons, that is, between the OA and control groups for females [p = 0.00018 ΔPAD (SE) = 2.99 (0.74) years] and for males [p = 6.6e-6 ΔPAD (SE) = 4.16 (0.86) years], as well as between the OA and CBP groups for males [p = 2.2e-5 ΔPAD (SE) = 4.69 (1.02) years] and this time also for females [p = 0.0081 ΔPAD (SE) = 2.88 (0.95) years].

3.5. Associations between brain-PAD derived by DeepBrainNet with measures of pain and function

We found no significant variable of interest-by-sex interaction (H_var-by-sex) in the model brain-PAD ~ variable of interest * sex + scanner + age + (variable of interest | scanner), for any clinical pain, QST or function variable of interest. We found however, significant associations between brain-PAD and several clinical pain, QST and function variables (H_var) in the model without interaction, brain-PAD ~ variable of interest + sex + scanner + age + (variable of interest | scanner), which are depicted in Figures 4 and 5. Effect sizes and p-values for these associations are summarized in Tables 5 and 6. This was replicated after removing outliers (see Figures S1 and S2, and Tables S2 and S3 of the Supplemental Materials). In all linear models, rejection of composite normality of the residuals failed after correcting for multiple comparisons (p > 0.05).

Figure 4. Linear mixed regression on clinical pain variables in the MSK pain only sample.

Significant (corrected p < 0.05) associations between variables characterizing clinical pain and the adjusted normalized brain-PAD are shown, restricted to the participants having musculoskeletal (MSK) pain. P-values (indicated in the title of each subplot) were corrected using FDR correction across the 13 clinical pain variables tested. Pairwise deletion was used for missing data. The resulting Degrees of Freedom (DoF) is specified in the title of each subplot. DBN: DeepBrainNet.

Figure 5. Linear mixed regression on function variables in the whole Sample.

Significant (corrected p-value < 0.05) associations between variables characterizing experimental pain and function and the adjusted normalized brain-PAD are shown (values adjusted for effects of all independent variables in the mixed model brain-PAD ~ variable of interest + scanner + sex + age + (variable of interest | scanner) except variable of interest). P-values (indicated in the title of each subplot) were corrected using FDR correction across the 30 variables tested. Pairwise deletion was used for missing data. DoF: Degrees of Freedom. DBN: DeepBrainNet.

Table 5.

Results of the linear mixed model regression of the brain-PAD on the variables characterizing clinical pain, restricted to those participants having musculoskeletal (MSK) pain.

Variable	Cohen’s f2	FDR corrected P-value	Bonferroni corrected P-value	DoF
GCPS-Pain Intensity	0.08m	0.016		120
WOMAC-Pain	0.05s	0.029		120
SF-MPQ-2-Continuous	0.15m	0.0022	0.0024	117
SF-MPQ-2-Intermittent	0.07m	0.017		116
SF-MPQ-2-Neuropathic	0.05s	0.029		118
SF-MPQ-2-Affective	0.07m	0.005	0.015	166
SF-MPQ-2-Total	0.13m	0.0022	0.0044	115
KL Index	0.10m	0.014		100
Pain Length	0.02s	0.036		233
Pain Duration	0.05s	0.024		156

Note. P-values were corrected for multiple comparisons (using FDR and Bonferroni corrections) across all 13 clinical pain variables. Not significant corrected P-values (p > 0.05) are not shown. Superscripts “s”, “m” and “l” indicate small (0.1² ≤ f² < 0.25²) and medium (0.25² ≤ f² < 0.4²) and “large” (f² > 0.4²) effect sizes, respectively. DoF: Degrees of Freedom.

Table 6.

Results of the linear mixed model regression of the brain-PAD on the variables characterizing experimental pain and function using the whole sample.

Variable	Cohen’s f2	FDR corrected P-value	Bonferroni corrected P-value	DoF
CSQ-R-Catastrophizing	0.08m	0.0011	0.0033	217
CSQ-R-Passive Coping	0.05s	0.029		217
PANAS-Negative Affect	0.08m	3.2e-05	6.3e-05	306
Somatization	0.08m	0.0015	0.006	194
PROMIS-Anxiety	0.07m	0.0028	0.014	197
PROMIS-Depression	0.06s	0.006	0.036	198
PROMIS-Sleep	0.04s	0.036		191
Severity of Insomnia	0.05s	0.012		187
SPPB Total Score	0.13m	2.1e-06	2.1e-06	265

Note. P-values were corrected for multiple comparisons (using FDR and Bonferroni corrections) across all 31 variables tested. Not significant corrected P-values (p > 0.05) are not shown. Superscripts “s”, “m” and “l” indicate small (0.1² ≤ f² < 0.25²) and medium (0.25² ≤ f² < 0.4²) and “large” (f² > 0.4²) effect sizes, respectively. DoF: Degrees of Freedom.

Tables 7 and 8 also report the partial correlations between each variable of interest and brain-PAD for each sex and their comparison using the z-test for the comparison between correlations. Both tables show significant correlations for a subset of the variables in Tables 5 (and Figure 4) and 6 (and Figure 5), respectively, with the same sign of the associations.

Table 7.

Partial regressions between variables characterizing clinical pain and brain-PAD for each sex and their comparison, restricted to those participants having musculoskeletal (MSK) pain.

Variable	DoF: correlation (p-value)			P-value of sex difference
	Females	Males	Both
GCPS-Pain Intensity	66: 0.28 (0.022)	32: 0.21 (0.47)	113: 0.22 (0.027)	0.77
WOMAC-Pain	66: 0.25 (0.032)	32: 0.17 (0.57)	113: 0.2 (0.038)	0.77
SF-MPQ-2-Continuous	83: 0.23 (0.032)	52: −0.081 (0.66)	150: 0.12 (0.16)	0.42
SF-MPQ-2-Intermittent	64: 0.42 (0.00075)	31: 0.21 (0.47)	110: 0.31 (0.004)	0.57
SF-MPQ-2-Neuropathic	63: 0.31 (0.015)	31: 0.22 (0.47)	109: 0.27 (0.0074)	0.77
SF-MPQ-2-Affective	65: 0.3 (0.015)	31: 0.11 (0.66)	111: 0.21 (0.035)	0.57
SF-MPQ-2-Total	89: 0.39 (0.00075)	56: 0.057 (0.66)	160: 0.24 (0.005)	0.42
KL Index	63: 0.41 (0.0009)	30: 0.22 (0.47)	108: 0.31 (0.004)	0.57

Note. P-values were corrected for multiple comparisons using FDR correction across all 13 variables tested. Significant correlations in bold font. DoF: Degrees of Freedom.

Table 8.

Partial regressions between variables characterizing experimental pain and function and brain-PAD for each sex and their comparison, for the whole sample.

Variable	DoF: correlation (p-value)			P-value of sex difference
	Females	Males	Both
Pressure Index	155: −0.17 (0.1)	80: −0.12 (0.53)	250: −0.2 (0.005)	0.96
CSQ-R-Catastrophizing	130: 0.29 (0.0052)	66: 0.18 (0.43)	211: 0.25 (0.0016)	0.96
PANAS-Negative Affect	178: 0.28 (0.0031)	109: 0.24 (0.1)	302: 0.25 (9.6e-05)	0.96
Somatization	118: 0.31 (0.0031)	52: 0.16 (0.53)	185: 0.22 (0.0097)	0.96
PROMIS-Anxiety	118: 0.26 (0.02)	55: 0.23 (0.36)	188: 0.24 (0.005)	0.96
PROMIS-Depression	119: 0.25 (0.021)	55: 0.19 (0.43)	189: 0.23 (0.005)	0.96
Severity of Insomnia	113: 0.22 (0.057)	50: 0.21 (0.43)	178: 0.21 (0.014)	0.97
SPPB Total Score	161: −0.28 (0.0031)	84: −0.32 (0.048)	260: −0.29 (5.2e-05)	0.96

Note. P-values were corrected for multiple comparisons using FDR correction across all 31 variables tested. Significant correlations in bold font. DoF: Degrees of Freedom.

4. Discussion

This is the first investigation on how brain age, predicted by a CNN method (DeepBrainNet), relates to MSK pain, using MRI scans from different cohorts, scanners, ages (i.e., 19–83), and MSK pain types (i.e., OA and CBP pain). To our knowledge, this is the largest and most heterogeneous sample ever used to relate brain aging to chronic pain. As a consequence of this heterogeneity, we found a significant variability in brain age prediction across MRI scanners that, if not accounted for, would have hindered the ability to detect chronic pain-brain aging associations.

Irrespective of sex, individuals with chronic OA pain had about 3–4.7 years “older” appearing brains compared to controls and CBP. This aligns with our previous work where older individuals with chronic pain had older brains compared to matched healthy controls [20]. They are also partially in agreement with the reports by Hung et al. (2022) [36]. Like in their work (we refer to their Figure 2A), we found that OA, but not CBP individuals, had significantly older appearing brains compared to controls. We also replicated their result that OA participants had older appearing brains than controls for each sex (their Figure 2B). However, we did not replicate their significant (though weak) difference between CBP and controls in females (their Figure 2B). This could owe to several methodological differences, such as differences in sample size and etiologies within the back pain groups. Note that despite having selected their OA and back pain participants from different studies, their controls for both conditions came from a third common study, which possibly precluded controlling for the effect of study, rendering difficult to determine whether their observed differences were pain-related or scanner-related. Also, Hung et al. (2022) assessed group differences separately for each sex, which does not allow for the determination of sex-related effects on the group differences. We explicitly answered this question for the first time by testing the actual sex-differences in these differential effects of chronic pain via the moderation analysis and found no significant results.

The fact that accelerated brain age seem to only manifest in OA is intriguing. However, it is not surprising that brain age happens to be different between OA and CBP pains since their brain morphological signatures have been reported to be significantly different [3]. Moreover, even though brain structure is significantly different between CBP and healthy controls [3], they could possibly have similar brain age estimates given the many-to-one nature of the MRI-to-brain age map. Novel spatially distributed brain age prediction methods, e.g., that based on the U-Net architecture [48], could help to determine more specific spatial brain age signatures of different MSK pain. On the other hand, the lack of sex-related effects is unexpected, given the increased pain sensitivity and risk for clinical pain commonly being observed among women attributed to a variety of mechanisms [4,27,32]. Again, sex-related differences in accelerated brain aging might only be observable at the local level, as recent reports suggest [52,63].

Contrary to a previous report by Sörös and Bantel (2020) [54], we did find a significant difference between all MSK pain participants and controls. Hung et al. (2022) [36] had hypothesized that the lack of a significant MSK pain-control difference reported by Sörös and Bantel (2020) [54] owed to the fact that this difference differed among pain types and thus pain effects could not be detected by merging all MSK pain participants. Our differential MSK-control differences among pain types is evidence for this hypothesis, but our bigger sample size could have also allowed brain age values from the OA group to drive towards a significant MSK-control difference.

We also found that older appearing brains were associated with greater intensity of pain, greater severity of the sensory (continuous, intermittent and neuropathic) and affective dimensions of pain, greater pain-related interference with daily activities, and greater radiographic severity of knee joint pathology in participants with knee OA pain. Overall, this suggest that accelerated brain aging could be linked to structural aberrations associated with pain severity. Moreover, after controlling for age, older appearing brains were associated with shorter pain durations, replicating the result by Hung et al., 2022 using the smaller OpenPain BNCM sample [36]. This association suggests that provided that pain severity is accounted for, MSK groups with longer pain durations will have younger appearing brains, which will thus differ less from those of control groups. This further supports the plausibility of the significant difference in brain aging between MSK pain and controls in older adults found by our group [20] that could not be replicated by Sörös and Bantel (2020) [54], since pain durations were shorter in the former sample compared to the latter (6.3 ± 8.8 years versus 15.9 ± 11 years, respectively), even though pain severity in both samples was similar (5.2 ± 1.9 versus 5 ± 2, respectively). Although seemingly counterintuitive, a negative pain duration-brain aging association after controlling for age may reflect that those individuals of a given age with more recent chronic pain onsets (that were older at onset) suffer in ways that are associated with more deleterious brain changes compared to those with longer pains (that were younger at onset). Figure S5 illustrates this. Recently, we have also reported this type of behavior between pain durations and brain functional connectivity [61].

Accelerated brain aging was not generally significantly associated with any experimental pain measure, suggesting that brain alterations associated with accelerated brain aging might not occur in primary areas implicated in experimental pain sensitivity. Conversely, older appearing brains were associated with greater pain catastrophizing, passive coping, negative affect, depressive symptomology, anxiety, sleep impairments, severity of insomnia and worse physical function, suggesting that accelerated brain aging could owe to alterations in areas implicated in the person’s general functioning, in tandem with his/her clinical pain characterization. These results also resemble others in the literature. For example, using a different brain age predicting method based on GPR [16], we reported that older appearing brains were associated with lower positive affect in older adults [20] and higher negative affect, more in vivo coping strategies and pain catastrophizing [37]. This confirms that the results are consistent across different brain age methodologies.

The association between brain age and either clinical pain or function was not moderated by sex. Provided that brain aging is a proxy of health outcomes, this aligns with a report that sex did not moderate the association between pain (or pain-related health outcomes) and several psychological factors in a large sample of chronic pain participants [55]. On the other hand, the correlations between brain age and either clinical pain or function were only significant for females. However, since none of these correlations significantly differed among sexes, we have no concluding evidence of sex effects on these associations, as these differences in significance could simply owe to the fact the subsamples of females were larger than those of males and thus more powered. More sex-balanced samples should be used to explore these differences.

Our study has several limitations. First, since this is a cross-sectional study, we are only testing associations and not causal relationships. Also, most measures were only available in the UF/UAB studies. This consistent lack of measures limits controlling for the effect of important confounders like depressive symptoms, medications, etc. Additionally, the age distribution in the OA group significantly differed from the controls and CBP group, owing to OA predominantly manifesting in middle and older ages. However, this should not be of concern. First, our independent variable is age-independent: theoretically, because brain-PAD is an age-independent deviation from chronological age by design; and in practice, because we removed the effects of chronological age by including it as a covariate. Second, by including scanner as a covariate, each pain group is compared with its own age-matched control group in a pooled error model. Nevertheless, to address doubts about a possible age-related sample selection bias, we repeated the analysis only using 44+ year-old participants, and found similar results (see Figure S3)—coincidently, this subsampling also eliminated the sex-related unbalance in age distributions, also helping to clear out any additional concern related to this issue (see Figure S4). Finally, we did not investigate which specific brain areas may be experiencing accelerated aging. Future measures of local contributions to brain age (e.g., using explanation maps [63]) may be more sensitive to different chronic pain conditions.

Conclusions

Using DeepBrainNet to predict brain age, and a large multi-center sample, we found that OA participants have older appearing brains compared to controls and CBP participants, whereas no significant difference was observed between CBP participants and controls. We also found significant associations between the predicted brain age difference and several measures of severity and comorbidities of chronic pain. Our results hint that more sophisticated MRI-based brain-age algorithms (e.g., local brain age predictions) may provide simplistic, clinically accessible and easily implementable biomarkers of chronic pain. Since several modifiable lifestyle risk factors [9,29], including body-mass-index, waist-to-hip ratio smoking and drinking [46], may be related to brain age, these biomarkers may also be used to monitor treatment outcomes.