Are magnetic resonance imaging features associated with intermittent and constant pain in knee osteoarthritis? A cross-sectional study

Abstract

Objective

This study aims to explore the relationship between MRI structural morphotypes and intermittent and constant pain in knee osteoarthritis (KOA).

Method

This cross-sectional study utilised data from the osteoarthritis initiative 48-month visit, including 6631 knees from 3494 individuals. Pain was assessed using the ICOAP Score and categorized into four types (no pain (NP), constant pain (CP), intermittent (IP), both intermittent and constant pain (IACP)). MRI images were acquired using 3T scanner and assessed by ROAMES. MRNet CNN and deep learning methods were used to calculate the probability of bone, meniscus/cartilage, inflammatory and hypertrophic structural morphotypes. The predictive model was based on the best subset algorithm.

Result

Structural morphotypes, T2 relaxation times and cartilage thickness varied significantly among the NP, IP, CP, and IACP groups. Structural phenotypes probabilities generally increased from NP to IP to CP, being highest in IACP group. The inflammatory morphotype was associated with all pain subgroups, the bone morphotype was mainly associated with IP, the meniscus/cartilage morphotype was associated with IP and IACP, and the hypertrophic morphotype was more strongly associated with CP than IP. ROC analyses showed modest discriminatory performance, with AUCs up to 0.70 for bone and 0.73 for hypertrophic morphotype probability.

Conclusion

Pain types in OA were associated with MRI structural morphotypes. The probability of the hypertrophic morphotype can be used to distinguish between IP and CP subgroups. A predictive model that incorporates the bone morphotype, inflammatory morphotype, and hypertrophic morphotype can effectively predict the presence or absence of pain.

Graphical abstract

1. Introduction

Pain major cause of disability in knee osteoarthritis (KOA) and it is the commonest presentation. The pain experience in KOA a complex multifaceted phenomenon, which remains poorly understood [1]. The situation is compounded by the fact that some patients with KOA experience intermittent or constant pain (CP), while others experience both types of pain [2]. There is interest in whether these pain types can be predicted or explained by previously identified structural KOA phenotypes, as this knowledge could improve understanding of the mechanisms underlying pain patterns heterogeneity in KOA. Establishing such associations may improve understanding of pain heterogeneity and support more personalized approaches to KOA management.

In KOA, abnormal joint loading and impaired tissue adaptation contribute to changes in articular cartilage, subchondral bone, and synovium. These changes include cartilage matrix degradation, dysregulated repair, synovial inflammation, and increased subchondral bone turnover. Recognition of these tissue-level changes has led to the classification of MRI -defined structural phenotypes. An OA phenotype has been defined as “a type of OA which shares distinct underlying pathobiological and pain mechanisms with their structural and functional consequences, including the severity of pain” [3]. MRI is used extensively to characterize structural disease severity in KOA based on tissue changes [4,5].

Proposed structural phenotypes proposed include the meniscus/cartilage, inflammatory, atrophic, hypertrophic and subchondral bone phenotypes [4,6] (see Fig. 1). The meniscus-cartilage phenotype has substantial meniscal/cartilage damage, whereas the subchondral bone phenotype has widespread bone marrow lesions. The inflammatory phenotype is characterized by joint effusion or synovitis. Hypertrophic and atrophic phenotypes are distinguished by prominent osteophytes or minimal osteophyte formation, respectively [7]. However, structural phenotypic characterization is still developing, and there is considerable variation in how to operationalize the concept of structural OA phenotypes.

Fig. 1.

Representative figure of MRI structural phenotypes [16] (inflammatory (A), meniscus/cartilage(B), bone(C), hypertrophic(D).

It is increasingly apparent that these structural phenotypes are associated with KOA progression and that some MRI findings are correlated with knee pain [8]. However, few studies have examined whether MRI-defined structural phenotypes are differentially associated with different type of pain. This distinction is clinically relevant as intermittent pain is believed to be a feature of early KOA [9]. Identifying associations between structural phenotypes and pain patterns may improve understanding of KOA pain mechanisms and inform future phenotype-guided research.

This present study aimed to determine the association between MRI biomarkers and MRI-defined structural phenotypes and CP, intermittent (IP), both intermittent and constant pain (IACP), and no pain (NP) in participants with knee OA. In addition, we examined whether a particular MRI biomarker/phenotype was associated/predicted either IP or CP.

2. Methods

2.1. Participants

Data was obtained from the Osteoarthritis Initiative (OAI), a publicly available dataset. The OAI is a multicentre longitudinal cohort study of 4796 individuals with or at increased risk of KOA [10]. The Institutional Review Board at each of the sites approved the study, with written informed consent obtained from all participants. Briefly, OAI participants were aged 45–79 years and had, or were at risk of, symptomatic knee OA. Exclusion criteria included inflammatory arthritis, contraindications to 3T MRI, and bilateral end-stage knee OA requiring surgery. We performed knee-level analyses of the OAI data at the 48-month visit, the initial collection of the Intermittent and Constant Osteoarthritis Pain (ICOAP) measure, using data from the period that covered 6631 knees from 3494 individuals.

2.2. Assessment of outcomes

2.2.1. Knee radiograph and pain severity assessment

The severity of KOA on weight-bearing knee radiographs was assessed using the Kellgren-Lawrence grading score [11], and ≥2 was classified as radiological OA.

2.2.2. IACP assessment

IACP was evaluated via the measure of ICOAP, a widely used scale that has excellent reliability and validity [12]. ICOAP assesses pain over the previous 7 days using 11 items covering pain intensity, frequency, impact on sleep and quality of life, and emotional impact. The frequency of pain was also assessed, with the following response options: “never” (0), “rarely”, “sometimes”, “often”, and “very often” (4).

A raw score from 0 to 24 was converted to a final score from 0 to 100. Using these scores, participants were classified into four pain subgroups as follows: (1) NP (means 0 on both CP and IP), (2) IP (means 0 on CP and >0 on IP) (3) CP (means 0 on IP and >0 on CP), and (4) IACP (means > 0 on both CP and IP). Further, the participants completed the KOA Outcomes Score (KOOS) and Western Ontario and McMaster (WOMAC) questionnaires. Both these questionnaires were used to assess OA symptoms and functionality, with lower scores on the KOOS and higher scores on the WOMAC indicating more severe pain symptoms [13].

2.2.3. MRI imaging biomarkers assessment

2.2.3.1. Structural phenotype assessment

We used five structural phenotypes previously identified using the simplified MRI instrument, Rapid OsteoArthritis MRI Eligibility Score (ROAMES) [6]. This approach used MRI changes in the three compartments (medial tibia (MT), patellofemoral (PAT) and lateral tibia (LT)) of the knee to stratify knees to different structural phenotypes [6,14,15]. Rapid OsteoArthritis MRI Eligibility Score (ROAMES) grades Hoffa-synovitis, effusion-synovitis, meniscus/cartilage damage, bone marrow lesions, and osteophytes on ordinal scales derived from MOAKS. The grading of features was performed based on the maximum score within each knee or compartment for each feature, and these phenotypes were identified as follows [6]: the inflammatory phenotype was characterized by Hoffa- or effusion-synovitis; the bone phenotype by large bone marrow lesions; the hypertrophic phenotype by large osteophytes with limited cartilage damage; and the meniscus/cartilage phenotype by meniscal and cartilage damage. The atrophic phenotype was not included in the present analysis because of the small number of cases.

Because MOAKS readings were unavailable for all participants, ROAMES, a simplified MOAKS-based scoring system, was used. Radiologist-assigned MOAKS/ROAMES grades for bone, meniscus/cartilage, inflammation, and hypertrophy were used as reference labels to train MRNet convolutional neural network (CNN) phenotype classifiers. The radiologist-graded images were divided into training (70%), validation (10%), and test sets (20%) for each CNN phenotype classifier. Finally, the MRNet CNN method was employed to calculate the probabilities of the bone, hypertrophy, inflammation, and meniscus/cartilage phenotypes in coronal and sagittal knee MRI images of all participants in OAI. To ensure the reliability and reproducibility of phenotype classification, the phenotype labels used for CNN training were derived from standardized ROAMES/MOAKS-based MRI feature definitions and radiologist-assigned reference grades. The CNN classifiers were trained using these radiologist-derived labels as the ground truth and evaluated using separate validation and test sets. Therefore, the phenotype probabilities used in the present analysis were generated through a standardized algorithm-based approach rather than subjective classification by the present investigators. For a detailed description, see the pre-publication [16,17].

2.2.3.2. T2 relaxation time assessment

T2 relaxation times were used to assess biochemical and microstructural joint changes in the articular cartilage. Altered T2 relaxation times reflect changes in cartilage hydration and collagen matrix organization [18]. T2 maps were generated from sagittal 2D multi-slice multi-echo spin-echo sequences with noise correction. Then, the T2 map was segmented for five different compartments and the average T2 value was calculated for each compartment [[19], [20], [21]]. The T2 measurements were conducted by our co-author, Valentina Pedoia, who developed an automated analysis methodology to comprehensively analyze the T2 values of the entire OAI dataset, ensuring accuracy and consistency in the segmentation process [22].

2.2.3.3. Cartilage thickness assessment

Cartilage thickness was measured on knee MRI using a fully automated method. [23]. The relevant data have been previously published [15]. The cartilage segmentation was subdivided into LT, lateral femur (LF), PAT, MT, and medial femur (MF), centroid of medial femur (cMF), full medial femur (fMF), full lateral femur (fLF), other area of medial femur (oMF), other area of lateral femur (oLF). The distance values were sampled at each skeletal point and averaged across all points and slices to determine the average thickness.

2.3. Constructing predictive models using the optimal subset algorithm

The primary outcome modelled in our analysis was pain patterns. MRI features significantly associated with pain groups were considered for model building. Model selection was performed using the best subsets algorithm and logistic regression, with the final model chosen according to the lowest Bayesian Information Criterion. This approach was selected because previous simulation studies have shown that best subset selection can provide strong prediction accuracy compared with stepwise and regularization methods [24].

2.4. Statistical methods

Continuous variables were summarized as mean ± SD for normally distributed data or as median (IQR) for skewed data, while categorical variables were presented as frequencies and percentages. Between-group differences across pain subgroups were assessed using generalized estimating equations to account for the correlation between bilateral knees within the same participant. Depending on data distribution, Gaussian family with identity link was used for continuous variables.

CNN-derived probabilities of the bone, inflammatory, hypertrophic, and meniscus/cartilage MRI structural phenotypes were analyzed as continuous variables. These phenotype probabilities were compared among the four pain subgroups: NP, IP, CP, and IACP. Univariable and multivariable multinomial logistic regression analyses were performed to evaluate the associations between quantitative MRI biomarkers, including MRI structural phenotype probabilities, and pain subgroups. Multivariable models were adjusted for age, sex, race, and BMI. Robust standard errors were clustered on participant ID to account for within-subject correlation. Results were reported as relative risk ratios with 95% confidence intervals. Model performance was assessed using the area under the ROC curve (AUC) with bootstrap resampling. Statistical significance was defined as P < 0.05. Analyses were performed using Stata version 17.0.

3. Results

3.1. Demographic characteristics

We included 3494 individuals and 6631 knees for knee-level analysis after excluding knees with missing MRI values. Table S1 presents the participants’ characteristics, in which, 3577 knees had NP, 2585 had only IP, 225 had only CP, and 244 had both. The sex and disease duration were similar in all groups. Compared with NP, knees with pain, particularly CP or IACP, had a higher proportion of advanced radiographic OA (KLG 3–4), higher BMI, lower KOOS scores, and higher WOMAC scores (Table S1). There were corresponding trends from NP to IP, to CP, to IACP groups.

3.2. MRI biomarkers in different pain subgroups

Significant differences were found among all pain groups and the NP group in the probabilities of bone, hypertrophic, inflammatory, and meniscus/cartilage phenotypes, as determined by generalized estimating equations models (Wald χ² = 279.46, 328.85, 170.30, 169.92, respectively; all P < 0.001). The probabilities of these phenotypes showed an increasing trend across pain subgroups, with the highest values observed in participants experiencing IACP, followed by those with CP, IP, and the lowest in NP group. The T2 relaxation times assessed LF were higher in IP and CP group than NP groups (Wald χ² = 12.64, P = 0.0055).

Additionally, the four groups differed significantly in the analysis of cartilage thickness. Within the pain groups, cartilage thickness tended to be lower at the LT, cMF, and PAT, but higher at the fLF, oMF, and oLF (Table 1). There was a progressive increase in the probabilities of bone, hypertrophic, inflammatory, and meniscus/cartilage phenotypes occurring from IP to CP subgroups, with the highest probability in IACP group (F = 123.4, Fig. 2). Notably, the probability of hypertrophic phenotype distinguished between the IP and CP groups and was more strongly associated with CP (t = 5.658), suggesting that hypertrophic structural changes were closely related to CP than IP.

Table 1.

The MRI biomarkers of subjects in the different pain subgroups.

MRI biomarkers∗		Pain subgroup				Wald χ2	P
		No pain	Intermittent pain	Constant pain	With both
Probability of inflammatory phenotype media (IQR)		0.176 (0.055–0.956)	0.387 (0.085–4.533)	0.456 (0.081–7.637)	0.730 (0.125–15.148)	170.30	0.0001
Probability of bone phenotype media (IQR)		2.151 (0.269–26.149)	14.394 (1.071–79.724)	24.637 (1.786–87.974)	34.988 (3.050–97.245)	279.46	0.0001
Probability of meniscus/cartilage phenotype media (IQR)		0.285 (0.052–3.076)	0.951 (0.102–20.034)	1.170 (0.093–27.760)	2.623 (0.139–60.555)	169.92	0.0001
Probability of hypertrophic phenotype media (IQR)		3.089 (0.940–12.567)	9.297 (1.937–38.930)	23.437 (3.241–74.759)	20.516 (4.805–57.360)	328.85	0.0001
T2 relaxation time median (IQR)	Lateral femur (LF)	36.635 (35.159–38.457)	36.957 (35.450–38.726)	37.060 (35.589–38.892)	36.574 (35.369–38.707)	12.64	0.0055
	Lateral tibia (LT)	29.541 (28.088–31.293)	29.587 (28.027–31.281)	29.686 (27.914–31.784)	29.609 (28.195–31.846)	3.60	0.3076
	Medial femur (MF)	39.535 (37.703)	39.882 (38.090–41.557)	39.973 (37.892–41.377)	39.863 (37.682–41.865)	7.12	0.0682
	Medial tibia (MT)	31.398 (30.013–32.823)	31.535 (30.049–32.874)	31.609 (29.922–32.985)	31.970 (30.273–33.191)	5.12	0.1632
	Patella (PAT)	33.758 (31.719–36.017)	33.504 (31.542–35.833)	32.803 (30.478–35.780)	32.667 (30.931–35.364)	2.87	0.4128
Cartilage thickness media (IQR)	Centroid of lateral femur (cLF)	1.802 (1.637–1.977)	1.813 (1.643–2.006)	1.778 (1.607–1.959)	1.820 (1.642–1.987)	3.72	0.2930
	Lateral tibia (LT)	1.999 (1.781–2.217)	1.984 (1.737–2.208)	1.964 (1.703–2.210)	1.939 (1.732–2.221)	10.68	0.0136
	Centroid of medial femur (cMF)	1.800 (1.616–2.017)	1.780 (1.573–2.016)	1.692 (1.502–1.938)	1.740 (1.504–2.006)	26.92	0.0001
	Medial tibia (MT)	1.702 (1.536–1.873)	1.710 (1.518–1.897)	1.653 (1.443–1.849)	1.661 (1.493–1.865)	9.05	0.0286
	Patella (PAT)	2.141 (1.839–2.417)	2.050 (1.739–2.329)	2.008 (1.778–2.345)	2.000 (1.763–2.889)	48.88	0.0001
	Full medial femur (fMF)	1.922 (1.757–2.084)	1.929 (1.773–2.093)	1.885 (1.745–2.041)	1.912 (1.755–2.126)	4.50	0.2127
	Full lateral femur (fLF)	1.982 (1.809–2.155)	1.996 (1.834–2.179)	1.977 (1.813–2.125)	1.995 (1.839–2.172)	8.58	0.0354
	Other area of medial femur (oMF)	1.963 (1.799–2.118)	1.980 (1.828–2.136)	1.939 (1.821–2.088)	1.968 (1.826–2.159)	10.30	0.0162
	Other area of lateral femur (oLF)	2.029 (1.849–2.217)	2.053 (1.880–2.244)	2.030 (1.861–2.194)	2.057 (1.881–2.228)	10.88	0.0124

Media (IQR) = Median(Interquartile Range).

Method: ∗ generalized estimating equations (GEE).

Fig. 2.

The distribution of MRI biomarkers across different pain subgroups. bone_prob: the probability of bone; men_prob: the probability of meniscus/cartilage; inf_prob: the probability of inflammatory; hyp_prob: the probability of hypertrophic. The values were standardized as the mean divided by the standard deviation to allow comparison across phenotypes with different value ranges. The numbers on the concentric circles indicate the radial axis scale and are displayed to one decimal place; higher values indicate higher standardized relative phenotype values. Each colored line represents one pain subgroup.

3.3. Association analysis of MRI biomarkers with different pain subgroups

Univariate and multivariate logistic regression analysis, accounting for within-participant clustering, showed higher probabilities between all pain subgroups and the NP subgroup regarding the probabilities of belonging to the bone, meniscus/cartilage, inflammatory, and hypertrophic phenotypes (Table 2). Interestingly, similar findings were observed in both the multivariate multinomial regression analyses and the adjusted multivariate multinomial regression analyses (Table 3).

Table 2.

The association of each MRI features with pain subgroups by Univariable multinomial logistic regression analysis.

qMRI biomarkers		Pain subgroup (Risk ratio (95%CI)P)
		No pain	Intermittent pain	Constant pain	With both
Probability of inflammatory phenotype		Ref	1.019 (1.016–1.022) 0.000	1.025 (1.019–1.031)0.000	1.030 (1.025–1.035)0.000
Probability of bone phenotype		Ref	1.012 (1.010–1.013)0.000	1.014 (1.011–1.018)0.000	1.017 (1.014–1.021)0.000
Probability of meniscus/cartilage phenotype		Ref	1.012 (1.010–1.014) 0.000	1.015 (1.010–1.019)0.000	1.019 (1.015–1.023)0.000
Probability of hypertrophic phenotype		Ref	1.018 (1.016–1.020) 0.000	1.029 (1.024–1.033)0.000	1.026 (1.022–1.030)0.000
T2 relaxation time (ms)	Lateral femur (LF)	Ref	1.051 (1.023–1.081) 0.000	1.011 (0.942–1.086)0.756	1.034 (0.961–1.114)0.373
	Lateral tibia (LT)	Ref	1.002 (0.974–1.031) 0.899	1.044 (0.972–1.121)0.235	1.054 (0.996–1.116)0.070
	Medial femur (MF)	Ref	1.035 (1.008–1.063) 0.010	1.032 (0.956–1.114)0.421	1.038 (0.960–1.123)0.348
	Medial tibia (MT)	Ref	1.025 (0.991–1.059) 0.147	1.051 (0.958–1.153)0.293	1.082 (0.993–1.180)0.072
	Patella (PAT)	Ref	0.995 (0.977–1.013)0.559	0.956 (0.903–1.013)0.125	0.999 (0.932–1.072)0.991
Cartilage thickness (mm)	Centroid of lateral femur (cLF)	Ref	1.205 (0.969–1.499)0.094	0.826 (0.450–1.514)0.536	1.059 (0.581–1.930)0.852
	Lateral tibia (LT)	Ref	0.793 (0.667–0.944)0.009	0.665 (0.420–1.052)0.081	0.625 (0.392–0.995)0.048
	Centroid of medial femur (cMF)	Ref	0.738 (0.613–0.889)0.001	0.368 (0.229–0.592)0.000	0.521 (0.302–0.898)0.019
	Medial tibia (MT)	Ref	1.107 (0.885–1.384)0.372	0.496 (0.270–0.912)0.024	0.646 (0.339–1.231)0.184
	Patella (PAT)	Ref	0.624 (0.540–0.721)0.000	0.578 (0.409–0.819)0.002	0.544 (0.382–0.774)0.001
	Full medial femur (fMF)	Ref	1.185 (0.923–1.521)0.183	0.654 (0.365–1.206)0.174	1.034 (0.516–2.073)0.924
	Full lateral femur (fLF)	Ref	1.400 (1.111–1.765)0.004	0.995 (0.545–1.814)0.987	1.241 (0.688–2.234)0.472
	Other area of medial femur (oMF)	Ref	1.489 (1.155–1.920)0.002	0.977 (0.532–1.793)0.939	1.504 (0.745–3.034)0.255
	Other area of lateral femur (oLF)	Ref	1.446 (1.158–1.805)0.001	1.104 (0.624–1.950)0.735	1.312 (0.755–2.280)0.335

Method: ∗ Multinomial logistic regression with robust standard errors clustered on participant ID.

Table 3.

The association of each MRI features with pain subgroups by Multivariate multinomial logistic regression analysis.

qMRI biomarkers		Multivariate multinomial logistic regression analysis (Risk ratio (95%CI)P)				Multivariate multinomial logistic regression analysis Adjusted by age, race and BMI
		No pain	Intermittent pain	Constant pain	With both	Intermittent pain	Constant pain	With both
Probability of inflammatory phenotype		Ref	1.006 (1.001–1.011) 0.021	1.011 (1.002–1.020) 0.022	1.013 (1.005–1.022) 0.002	1.007 (1.002–1.012) 0.007	1.015 (1.005–1.024) 0.002	1.016 (1.008–1.025) 0.000
Probability of bone phenotype		Ref	1.005 (1.003–1.008) 0.000	1.002 (0.996–1.008) 0.551	1.002 (0.996–1.008) 0.514	1.006 (1.003–1.008) 0.000	1.001 (0.995–1.007) 0.697	1.001 (0.996–1.007) 0.634
Probability of meniscus/cartilage phenotype		Ref	1.003 (1.000–1.007) 0.065	1.001 (0.993–1.009) 0.787	1.009 (1.002–1.017) 0.012	1.005 (1.000–1.008) 0.012	1.007 (0.999–1.015) 0.086	1.014 (1.007–1.022) 0.000
Probability of hypertrophic phenotype		Ref	1.008 (1.004–1.013) 0.000	1.028 (1.018–1.037) 0.000	1.011 (1.002–1.021) 0.031	1.006 (1.002–1.011) 0.005	1.018 (1.008–1.028) 0.000	1.004 (0.994–1.014) 0.440
T2 relaxation time (ms)	Lateral femur (LF)	Ref	1.023 (0.987–1.060) 0.214	0.959 (0.878–1.048) 0.358	0.994 (0.912–1.082) 0.884	1.015 (0.978–1.052) 0.432	0.965 (0.880–1.058) 0.448	0.995 (0.911–1.086) 0.911
	Medial femur (MF)	Ref	1.009 (0.975–1.044) 0.50	1.051 (0.954–1.158) 0.317	1.019 (0.930–1.117) 0.685	1.019 (0.984–1.055) 0.296	1.098 (0.993–1.215) 0.069	1.055 (0.959–1.161) 0.269
Cartilage thickness (mm)	Lateral tibia (LT)	Ref	0.905 (0.655–1.250) 0.545	1.346 (0.560–3.237) 0.596	1.461 (0.615–3.467) 0.390	0.767 (0.552–1.066) 0.115	1.350 (0.562–3.246) 0.502	1.329 (0.566–3.120) 0.514
	Centroid of medial femur (cMF)	Ref	0.590 (0.388–0.896) 0.013	0.633 (0.234–1.713) 0.68	0.492 (0.179–1.350) 0.168	0.627 (0.409–0.960) 0.032	0.564 (0.205–1.547) 0.266	0.467 (0.163–1.341) 0.157
	Medial tibia (MT)	Ref	0.946 (0.557–1.600) 0.836	0.312 (0.082–1.187) 0.088	0.289 (0.069–1.207) 0.089	0.816 (0.478–1.393) 0.456	0.533 (0.138–2.063) 0.362	0.391 (0.092–1.664) 0.204
	Patella (PAT)	Ref	0.878 (0.670–1.101) 0.259	1.427 (0.741–2.748) 0.288	0.921 (0.495–1.714) 0.796	0.771 (0.610–0.974) 0.029	0.906 (0.476–1.728) 0.765	0.645 (0.354–1.175) 0.152
	Full lateral femur (fLF)	Ref	0.384 (0.062–2.368) 0.302	0.042 (0.000–2.481) 0.127	0.114 (0.001–11.513) 0.356	0.609 (0.100–3.709) 0.590	0.140 (0.001–13.588) 0.400	0.344 (0.003–34.186) 0.649
	Other area of medial femur (oMF)	Ref	2.047 (1.114–3.759) 0.021	4.046 (0.915–17.893) 0.065	8.294 (1.741–39.528) 0.008	1.865 (1.005–3.462) 0.048	2.022 (0.440–9.316) 0.366	4.572 (0.924–22.610) 0.062
	Other area of lateral femur (oLF)	Ref	3.503 (0.676–18.163) 0.135	16.189 (0.384–682.500) 0.145	5.852 (0.111–308.955) 0.383	2.439 (0.476–12.494) 0.285	8.087 (0.131–499.973) 0.321	2.963 (0.058–151.227) 0.588

Method: ∗Multinomial logistic regression adjusted for age, race, and BMI, with robust standard errors clustered on participant ID.

In addition, multivariate multinomial regression analysis (adjusted with age, race, and BMI) results revealed higher probabilities of inflammatory phenotype and hypertrophic phenotype in the pain subgroups compared to the NP group. Inflammatory phenotype probability was associated with all pain subgroups compared with the NP group, (RRR = 1.007 (1.002–1.012) in IP; RRR = 1.015 (1.005–1.024) in CP; RRR = 1.016 (1.008–1.025) in IACP), bone phenotype probability was associated with IP group (RRR = 1.006 (1.003–1.008)), meniscus/cartilage phenotype probability was associated with IP group (RRR = 1.005 (1.000–1.008)) and IACP group (RRR = 1.014 (1.007–1.022)), hypertrophic phenotype probability was associated with IP group (RRR = 1.006 (1.002–1.011)) and CP group (RRR = 1.018 (1.008–1.028)). Overall, the inflammatory phenotype was associated with pain subgroups, the bone phenotype was mainly associated with IP, the meniscus/cartilage phenotype was associated with IP and IACP, and the hypertrophic phenotype was more representative of CP than IP.

3.4. MRI predictive biomarkers with pain

A prediction model was constructed based on the optimal subset algorithm, as can be seen in Table S2. Group 3, with the lowest Bayesian Information Criterion value, was the best prediction model and included the combination of bone, inflammatory, and hypertrophic phenotype probability variables. The best predictive model results showed that the probabilities of phenotypes were correlated with the occurrence of pain and that all of them could be used as predictors of pain (all P < 0.0001) (Table 4). The probabilities of bone, inflammatory and hypertrophic phenotypes were associated with the pain subgroups.

Table 4.

The Multinomial logistic regression of best model predictors in pain subgroups.

qMRI biomarkers	Pain subgroup (Risk ratio (95%CI)P)
	No pain	intermittent pain	constant pain	with both
Probability of bone phenotype	Ref	1.006 (1.005–1.008)0.000	1.003 (0.999–1.007)0.147	1.009 (1.005–1.014)0.000
Probability of inflammatory phenotype	Ref	1.010 (1.006–1.013)0.000	1.010 (1.004–1.017)0.001	1.018 (1.012–1.024)0.000
Probability of hypertrophic phenotype	Ref	1.011 (1.008–1.013)0.000	1.024 (1.018–1.029)0.000	1.014 (1.008–1.020)0.000

Method: ∗Multinomial logistic regression with robust standard errors clustered on participant ID.

Further, we used ROC curve analysis to demonstrate the ability of MRI phenotype to distinguish between NP and pain group. As shown in Table S3, for the probability of inflammatory phenotype, the AUC was 0.59 (95% CI, 0.579–0.608) in IP, 0.60 (95% CI, 0.558–0.642) in CP, 0.659 (95% CI, 0.620–0.698) for IACP, respectively (Fig. 3a). For the probability of bone phenotype, the AUC was 0.63 (95% CI, 0.617–0.645) in IP, 0.67 (95% CI, 0.635–0.706) in CP, 0.70 (95% CI, 0.661–0.731) for IACP, respectively (Fig. 3b). For the probability of meniscus/cartilage phenotype, the AUC was 0.60 (95% CI, 0.585–0.614)) in IP, 0.60 (95% CI, 0.557–0.640) in CP, 0.66 (95% CI, 0.619–0.695) for IACP respectively (Fig. 3c). For the probability of hypertrophic phenotype, the AUC was 0.63 (95% CI, 0.620–0.648), in IP, 0.72 (95% CI, 0.684–0.757) in CP, 0.73 (95% CI, 0.700–0.763) for IACP, respectively (Fig. 3d). Overall, these findings suggest that MRI structural phenotype probabilities may provide complementary information for distinguishing pain phenotypes in KOA; however, their standalone predictive utility appears limited.

Fig. 3.

The ROC curve of predictors in different pain groups (Compared with no pain). (a) The ROC curves of the probability of inflammatory phenotype; (b) The ROC curves of the probability of Bone phenotype; (c) The ROC curves of the probability of meniscus/cartilage phenotype; (d) The ROC curves of the probability of hypertrophic phenotype.

As shown in Table S4, there were increased odds of having a higher KL grade with IP, with CP, as well as IACP of subgroups compared to the NP group. The more severe the pain, the higher OR for the same KL grade, and at KL = 4, the risk of CP was higher than intermittent pain (OR (95%CI): 2.127 (1.295–3.494)).

4. Discussion

Our study demonstrated that previously defined MRI phenotypes, namely bone, hypertrophic, inflammatory and meniscal/cartilage phenotypes, were associated with different pain subgroups in KOA. In addition, we identified that different structural phenotypes were associated with varying pain subgroups and may provide complementary information for distinguishing pain patterns. These findings suggest that MRI structural phenotypes may help characterize pain heterogeneity in KOA; however, given the modest magnitude of some associations, these phenotypes should be interpreted as supportive information rather than stand-alone clinical predictors.

The first finding was that probabilities of bone, hypertrophic, inflammatory, and meniscus/cartilage phenotypes of MRI were significantly different in pain subgroups (IP, CP and IACP) and the NP group and increased incrementally from IP to CP subgroups and were highest in the IACP group. These findings are consistent with previous evidence that MRI-detected structural abnormalities are related to KOA symptoms, although our study further extends this by focusing on IP and CP patterns [25]. The associations between MRI structural phenotypes and pain patterns were phenotype-specific. In the adjusted model, the bone phenotype was significantly associated with IP but not CP, whereas the hypertrophic phenotype showed a stronger association with CP than IP. The inflammatory phenotype was associated with all pain subgroups, suggesting a broader relationship between synovitis-related inflammatory changes and KOA pain. The meniscus/cartilage phenotype was associated with IP and IACP, but not CP, indicating that meniscal and cartilage abnormalities may be more relevant to episodic or mixed pain patterns. These findings suggest that different structural phenotypes may be differentially related to different pain patterns in KOA. These findings may have potential clinical implications for phenotype-guided KOA pain management [26]. Inflammatory phenotypes may be relevant to synovitis- or inflammation-targeted approaches, whereas bone and meniscus/cartilage phenotypes associated with IP may suggest mechanically triggered or load-related pain, for which rehabilitation, biomechanical interventions, weight management, and activity modification may be relevant. The stronger association between the hypertrophic phenotype and CP may suggest a more persistent pain state requiring multimodal management. However, because treatment response was not evaluated, these interpretations should be considered hypothesis-generating rather than direct therapeutic recommendations.

Bone marrow lesions are believed to reflect increased pressure and altered force on the subchondral bone. Previous studies have linked the bone phenotype to radiographic and pain progression in KOA [27], and fluctuations in knee pain have been associated with bone marrow lesions [28]. The inflammatory phenotype was characterized by effusion-synovitis and Hoffa-synovitis, which may reflect increased cytokines, chemokines, and inflammatory molecules in synovial fluid and joint tissues [29,30]. Our finding is also consistent with previous study that synovitis was associated with CP in knee OA [31], and with studies linking inflammatory markers, effusion-synovitis, and marrow lesions to knee pain patterns [29,32]. The hypertrophic phenotype is characterized by osteophytes, bone overgrowth, and new bone formation, and osteophyte size has been linked to KOA progression, cartilage damage, and knee symptoms [33,34]. The meniscus/cartilage phenotype was associated with IP and IACP, supporting a potential role of meniscal and cartilage abnormalities in episodic or mixed pain patterns [35,36]. The present results first confirm the relationship between meniscal phenotype and pain and provide evidence to investigate the mechanisms of OA further.

The hypertrophic phenotype was more strongly associated with CP than IP, suggesting that hypertrophic structural changes may be more representative of CP. This is consistent with evidence that hypertrophic phenotypes and larger osteophytes are related to joint space narrowing, cartilage damage, and more advanced structural disease [37,38]. In our study, advanced radiographic KOA was also more strongly associated with CP than IP. A model containing bone, inflammatory, and hypertrophic phenotypes showed discriminatory ability for pain presence, although its predictive utility should be interpreted cautiously.

In addition, our study demonstrated that previously identified risk factors for KOA pain were associated with these IP, CP and IACP groups. The four subgroups differed significantly in terms of age, BMI, race, and family history. Although many studies have shown that KOA is associated with age, our findings objectively demonstrate that age is related to differences among KOA pain subgroups. Given that age may also reflect disease duration or progression, this potential relationship merits further exploration in future studies. Previous research has shown that has shown that individuals with increased BMI are more likely to have knee pain than persons with normal weight [39]. Our study showed that all three pain subgroups are more prone to have increased BMI. Other researchers have observed differences in pain severity, self-reported disability, and reduced function in different racial groups [40]. Similarly, individuals with a family history of KOA face a higher risk of worsening knee pain, regardless of structural damage and changes [41]. Our study corroborates both these findings by showing that non-white races and persons with a family history of arthritis were more report CP (or IACP) than the NP and IP subgroups. Unsurprisingly, the KOOS and WOMAC scores are significantly different in the pain subgroups, and both these scales are heavily influenced by self-reported pain [42,43]. More severe pain was associated with higher Kellgren-Lawrence grading. At advanced radiographic KOA, the risk of CP was higher than IP. These findings suggest that pain patterns may vary according to disease stage. IP may be more closely related to episodic or mechanically triggered structural changes, whereas CP and combined IP and CP may reflect greater cumulative structural burden or a more persistent pain state. Nevertheless, the modest magnitude of several associations suggests that these MRI structural phenotypes should not be interpreted as stand-alone clinical predictors at the individual-patient level. Instead, they may provide complementary information for understanding pain heterogeneity and disease-stage-related pain patterns in KOA.

T2 relaxation times are the sensitive marker of compositional changes in the cartilage matrix, capturing early biochemical alterations that may occur before structural degeneration becomes evident (including cartilage hydration, proteoglycan loss, and disruption of collagen) [44]. Cartilage thickness reflects gross morphological alterations, which are more indicative of advanced disease stages characterized by significant structural degeneration [45]. Our study found that the cMF of T2 relaxation times and the PAT of cartilage thickness were associated with IP in adjusted multivariable logistic regression, indicating an IP-specific signal. Because T2 primarily reflects early cartilage matrix alterations (e.g., collagen disorganization, increased water content), these findings suggest that episodic symptoms may emerge when matrix integrity is perturbed but overt morphologic loss is limited. By contrast, CP is more plausibly linked to greater structural burden (e.g., hypertrophic, inflammatory, and bone morphotypes), which early matrix-sensitive measures are less likely to discriminate. This interpretation is consistent with previous studies [21,46], which reported associations between pain and quantitative MRI measures, including T2 values and cartilage thickness, using OAI data, and supports a morphology-informed pain phenotyping framework in KOA.

Despite the importance of the findings above, this study has certain limitations. Firstly, the phenotypes used in this study were empirically defined based on a priori hypothesis. It is very likely that these phenotypes overlap and that multiple phenotypes exist in one person. Further, the discordance between structure and pain perception, has been previously established [47]. This may explain the relatively low AUC for predictive ability of MRI phenotype to distinguish between NP and the pain groups. Further, this phenomenon may explain why a combination of MRI phenotypes was more predictive of the pain types than one phenotype alone. As this study used an already existing dataset, this cohort had higher numbers of persons with IP when compared to CP and IACP. It is possible that a better strength of association may have been demonstrated if higher numbers of persons with CP and both IACP were included in this study. Due to the lack of detailed information on participants’ prior knee injuries or surgeries, we were unable to control for this potential confounder in our analysis. Similarly, physical activity was not adjusted for in the multivariable models because the available physical activity measures were not sufficiently complete and/or temporally aligned with the MRI and pain assessments in our analytical sample; therefore, residual confounding by physical activity cannot be excluded. The estimates were generally similar between the crude and adjusted multinomial logistic regression models, suggesting that the included covariates, including age, race, and BMI, did not substantially confound the observed associations. However, this should not be interpreted as evidence that all relevant confounding was fully controlled. Residual confounding from other unmeasured or incompletely measured clinical and behavioral factors cannot be excluded. Furthermore, this analysis is cross-sectional and cannot establish temporal sequence or causality. We are planning a prospective cohort with repeated MRI and standardized pain phenotyping to test temporality, validate the observed associations, and re-evaluate model performance.

The atrophic phenotype, defined as minimal osteophytes with severe cartilage damage, is the fifth ROAMES phenotype, which was not included due to the low number of cases. Due to the sample size, we did not perform a separate training and test set split, which may increase the risk of overfitting. In future work, we plan to incorporate k-fold cross-validation and external validation with independent datasets to improve model robustness and ensure generalizability. The phenotypes demonstrated relatively low predictive power in differentiating pain levels, with AUC values falling below the commonly accepted threshold for high predictive performance. This may be attributed to the high variability in individual pain experiences. Future research should increase the sample size and incorporate additional covariates, such as biochemical markers or psychosocial factors, to improve the model’s predictive performance.

5. Conclusions

This study has identified that MRI phenotypes in the pain groups were different from those demonstrated in the NP group. The probability of the hypertrophic morphotype distinguished intermittent from CP. Integrating bone, inflammatory, and hypertrophic morphotypes predicted pain presence and differentiated pain phenotypes. These structural features will help explain the pain experienced in KOA. Thus, the MRI-based structural associations in this study have the potential to be used to develop interventions for the clinical problem of KOA pain.

Author contributions

All authors contributed to the design of the research and interpretation of the data: DJH, DZ, and IA; conceptualization: DJH, DZ, and IA; analysis and interpretation of the data: VP, VV, DJH, and DZ; statistical expertise: VV and DZ; writing-original draft preparation: DZ and IA; writing-review and editing: LAD, XL, SQS and DJH. All authors gave feedback on drafts of this paper and read and approved the final manuscript.

Role of the funding source

This study was supported by the National Natural Science Foundation of China (82073494) and the NHMRC Investigator Grant (1194737); the Natural Science Basic Research Program of Shaanxi (2025JC-YBQN-1044); Shaanxi Postdoctoral Research Program(2024BSHSDZZ046); the Fundamental Research Funds for the Central Universities, Xi’an Jiaotong University (xzy012025154) and Postdoctoral Fellowship Program of China Postdoctoral Science Foundation (GZC20250513).

Handling Editor: Professor H Madry

Appendix A. Supplementary data

The following is the Supplementary data to this article: