MRI-based Extended Ordered Values More Efficiently Differentiate Cartilage Loss in Knees With and Without Joint Space Narrowing than Region-specific approaches using MRI or Radiography – Data from the OA Initiative

Abstract

Objective

The sensitivity to change of quantitative analysis of cartilage in knee osteoarthritis using MRI is compromised by the spatial heterogeneity of cartilage loss. We explore whether extended (medial-lateral) “ordered values” (OV) are superior to conventional approaches of analyzing subregional cartilage thickness loss and to radiography, in differentiating rates of progression in knees with and without joint space narrowing (JSN).

Methods

607 Osteoarthritis Initiative participants (308 without and 299 with baseline JSN at baseline) were studied over 12 months. Subregional femorotibial cartilage loss was determined in all knees, and changes in minimum joint space width (mJSW) in a subset of 290 knees. Subregional thickness changes in medial and lateral tibial and femoral cartilages were sorted in ascending order (OV1-16). A Wilcoxon rank-sum test was used to compare rates of change in knees with and without JSN.

Results

JSN-knees displayed greater cartilage loss than those without JSN, with minimal p-values of 0.008 for femorotibial subregions, 3.3×10⁻⁴ for medial OV1, and 5.4×10⁻⁷ for extended (medial and lateral) OV1. mJSW measurements (n=290) did not discriminate between longitudinal rates of change in JSN versus no-JSN knees (p=0.386), whereas medial OV1 (p=5.1×10⁻⁴) and extended OV1 did (p=2.1×10⁻⁵).

Conclusion

Extended OVs showed higher sensitivity to detecting differences in longitudinal rates of cartilage loss in knees with and without baseline JSN than anatomical (sub)regions and radiography. The OV technique also circumvents challenges of selecting particular regions “a priori” in clinical trials and may thus provide a powerful tool in studying risk factors or treatment efficacy in osteoarthritis.

INTRODUCTION

Quantitative magnetic resonance imaging (MRI) has emerged as a powerful tool for elucidating the natural progression and patho-physiology of osteoarthritis (OA), for identifying risk factors of OA, and for evaluating the effect of structure or disease modifying OA drugs (DMOADs)^1–4. However, recent studies employing MRI technology reported that longitudinal changes of cartilage thickness in osteoarthritis (OA) displayed a great amount of spatial heterogeneity between femorotibial joint compartments (medial, lateral), plates (tibia, femur), and subregions ^5–11. Previous studies hypothesized that cartilage loss in knee OA may preferentially occur in certain subregions of the femorotibial joint ^5,12, but recent evidence suggests that the MRI-based sensitivity to change for anatomically defined subregions is not relevantly improved when compared to the analysis of total cartilage plates: Although central femorotibial subregions generally displayed greater rates of change than peripheral ones ^5,9, the intersubject variability of central changes was also higher than for total cartilage plates ^6,9. Potential explanations for this observation are that only some knees show preferential central changes, and that once the cartilage is lost centrally, no further progression can be observed in central subregions. Moreover, a recent study showed that local meniscus lesions (in the anterior or posterior horn or body) are associated with higher rates of progression in immediately adjacent tibial cartilage subregions ¹³. The fact that meniscal lesions are frequent ¹⁴ and strongly related to OA progression ^15,16 provides a potential explanation, why rates of cartilage loss display strong spatial heterogeneity in peripheral subregions in OA.

As a potential solution to this challenge, Buck et al. ¹⁷ recently proposed a strategy for more efficiently measuring cartilage loss in OA by removing the link between magnitudes and locations of regional thickness changes in MRI. The authors showed that determining ordered values (OVs) of subregional change within the MEDIAL femorotibial compartment of each knee (medial OV approach) and then ranking the subregional change according to its magnitude, provided improved discrimination of cartilage loss between changes in healthy subjects and participants with MEDIAL radiographic OA. However, in general OA populations, a problem arises from the fact that some knees show preferential changes in the medial and others in the lateral femorotibial compartment, partly caused by differences in limb alignment ^16,18–21. In clinical trials, this can be circumvented by only selecting knees with either medial or lateral disease but this substantially increases the effort and cost involved in participant selection and also limits generalizability. Moreover, a recent study investigating the potential structure modifying effects of licofelone and naproxen ²² selected patients with MEDIAL femorotibial radiographic change and defined the MEDIAL compartment cartilage volume changes as the primary efficacy outcome measure. Although the primary outcome was reached in this study, the protective effect of licofelone was more evident in the lateral than in the medial compartment.

The objectives of the current study were:

1. to extend the proposed OV approach ¹⁷ to not only include medial but also lateral femorotibial subregions, in order to account for knees with both medial and lateral (radiographic) OA

2. to apply this approach to the analysis of cartilage thickness changes (i.e. cartilage loss) as a measure of OA progression in a large subset of knees with and without radiographic joint space narrowing (JSN) at baseline, provided by the OA Initiative (OAI) ^8,23–25.

3. to examine whether the extended OV approach shows a greater statistical sensitivity to differences in longitudinal cartilage changes between knees with and without baseline JSN than

   1. the medial OV approach,
   2. the region-based approach, and
   3. the minimal joint space width [mJSW] approach, using radiography.

4. to further explore the statistical specificity of the extended ordered values approach in relation to the region-based approach and the medial ordered values approach.

METHODS

Study participants

The study was based on the analysis of right knees from the OAI (public use data sets 0.2.2 [baseline clinical], 0.E.1 [baseline images], and 1.E.1 [12 month follow up images]) and was conducted in compliance with the ethical principles derived from the Declaration of Helsinki and in compliance with local Institutional Review Board, informed consent regulations, and International Conference on Harmonization Good Clinical Practices Guidelines. Knees were randomly selected based on a) ascending OAI ID and b) the “calculated” Kellgren and Lawrence grades (cKLG), derived from osteophyte and radiographic JSN readings performed at baseline at the OAI clinical sites, according to the OARSI atlas ²⁶. For the current study, we selected 607 knees with definite osteophytes and with either moderate JSN (299 knees with OARSI grade 1 or 2:113 male, 186 female) ²⁶ or without JSN at baseline (308 knees: 111 men, 197 female), because previous studies have shown a higher rate of cartilage thickness changes in knees with advanced radiographic OA (i.e. with baseline JSN) than in knees with less advanced radiographic OA (i.e. without baseline JSN) ^7,27–29. In this context it is worth noticing that although JSN is clearly associated with cartilage thickness loss ³⁰, there is substantial variability between OA participants, and JSN is additionally influenced by meniscus extrusion and degeneration ³¹.

For 290 of the above knees, longitudinal (quantitative) measurements of medial minimum radiographic joint space width (mJSW) ³² have recently been made available by the OAI (J Duryea, Brigham and Womens Hospital, Boston, MA) based on fixed flexion radiographs ³³ obtained at baseline and at 12 month follow-up. From these 147 displayed baseline JSN in site readings and 143 did not.

MR image analysis

Double oblique, coronal MR images were acquired at baseline and 12 month follow up, using a fast low angle shot sequence with water excitation (FLASHwe), 3 Tesla MR scanners (Siemens Magnetom Trio, Erlangen, Germany) and quadrature transmit-receive knee coils (USA Instruments, Aurora, OH); the imaging protocol and quality control procedures have been described in detail in previous publications ^8,23,24,34 (Figure 1). After a quality control step (M.H.) at the image analysis center (Chondrometrics GmbH, Ainring, Germany), the data were analyzed by seven readers, each with more than 3 years experience in cartilage segmentation. The segmentation was performed for paired baseline and 12 month follow-up images, the readers being blinded to the order of the acquisition as well as to the clinical and radiographic data. The subchondral bone area (tAB) and the cartilage surface area (AC) were traced manually in the medial (MT) and lateral tibia (LT) and in the weight-bearing central part of the medial (cMF) and lateral femoral condyle (cLF). All segmentations were quality controlled by an expert reader (S.M.) and were corrected by the readers, if necessary. The mean cartilage thickness over the total subchondral bone area, including denuded areas, (ThCtAB) was determined in cartilage plates (MT, LT, cMF and cLF) and compartments (medial femorotibial compartment =MFTC =MT+cMF and lateral compartment =LFTC =LT+cLF). Subregional thickness was determined in the central, external, internal, anterior, and posterior aspect of MT and LT, and in the central, external, and internal aspects of cMF and cLF, as described previously ¹² (Figure 1). The central subregions were set to cover 20% of the tAB in MT and LT, and 33% in cMF and cLF.

Figure 1.

Double oblique coronal fast low angle shot (FLASH) MR image with water excitation showing the regions of interest analyzed: MFTC = medial femorotibial compartment (=MT+cMF), LFTC = lateral femorotbial compartment (=LT+cLF); MT = medial tibia, cMF = central, weight-bearing part of the medial femoral condyle, LT = lateral tibia, cLF = central, weight-bearing part of the lateral femoral condyle.

The top part of the figures shows a reconstruction of the weight-bearing parts of the femoral condyles (cMF and cLF) and the lower part a reconstruction of the tibiae (MT and LT). (c|e|i|a|p = central|external|internal|anterior|posterior subregion of MT or LT. c|e|i = central|external|internal subregion of the central part of cMF or cLF.

Statistical analysis

Differences in subject characteristics between subjects with and without baseline JSN were assessed using two-sided t-tests. All longitudinal analyses including only MRI data were applied to the full cohort (n=607), whereas the comparison between longitudinal MRI and radiographic mJSW was performed for the subcohort (n=290). As a measure of progression, the mean change (MC) and the standard deviation (SD) of the change in µm for ThCtAB (MRI) and mJSW (radiography) between baseline and 12 month follow-up were determined. Percent changes were derived by relating the MC observed across a group to the respective average baseline value. The differences between the changes in the compared groups were described by the mean difference and 95% confidence intervals. The OV approach ¹⁷ was extended to comprise all 16 subregions in the femorotibial joint (5 in MT and LT, and 3 in cMF and cLF, respectively): Subregional changes (in ThCtAB) within each knee were sorted in ascending order, i.e. the subregion showing the most negative change (decrease in ThCtAB) was assigned to extended ordered value (eOV) 1, and the value of the subregion showing the smallest negative or greatest positive change (increase in ThCtAB) was assigned to eOV 16 (Figure 2).

Figure 2.

Graph showing the ordered values approach: A) The top spreadsheet shows the results (change in cartilage thickness [ThCtAB] in µm) in the femorotibial subregions (see Figure 1) of four example OAI subjects. B) The rates of change are ranked according to their magnitude in the middle spread sheet. C) The magnitudes of the changes (in µm) are then attributed to the orders in the bottom spread sheet.

The subregion with the most negative change (decrease in ThCtAB) in each subject is assigned to order 1, the subregion showing the second most negative change assigned to order 2, and the subregion showing the smallest negative or the greatest positive change (increase in ThCtAB) assigned to order 16. Note that differently located subregions contribute to order 1 in the four subjects shown.

To compare the rates of progression (cartilage thickness loss) in no-JSN and JSN knees, the MC and SD were evaluated for each compartment, cartilage plate and subregion as well as for the medial (mOV) and eOV approach (medial and lateral) (mOV 1 to 8, and eOV 1 to 16, respectively). The non-parametric Wilcoxon rank-sum test was used to determine whether the changes differed significantly between the JSN and the no-JSN knees, because the longitudinal changes may not be normally distributed.

Because the chance of at least one type I error increases with the number of parallel comparisons, and because the number of parallel comparisons differed between compartments (2 measures), cartilage plates (4 measures), subregions (16 measures), and ordered values (8 measures for medial OVs, 16 measures for extended OVs), the individual test significance levels were adjusted for the number of (parallel) comparisons (Bonferroni-Dunn correction for overall significance level = 0.05): p<0.025 for 2 compartments; p<0.0125 for 4 plates, p<0.003125 for 16 subregions, p<0.00625 for 8 medial OVs and p<0.003125 for 16 extended OVs. The significance levels were adjusted within each hierarchical category of joint compartments (n=2), cartilage plates (n=4) or cartilage subregions (n=16), but not across these categories, because lower hierarchical levels are contained in (and correlated with) higher levels.

To further explore the statistical sensitivity of the different MRI-measures, the bootstrapping approach ³⁵ was employed to simulate 10,000 “new” samples derived from the original study cohort by sampling the observed changes with replacement. The sample sizes (no-JSN and JSN) were kept constant.

The specificity of all measures is theoretically fixed during the testing procedure, as it corresponds to the level of false positives (significance level α), which is stated a priori. This is a theoretical assumption, however, and it is worthwhile to assess whether new testing procedures match the desired significance level. A randomization test assigning the observed changes in the JSN cohort for 10,000 times randomly without replacement to two subcohorts was employed for this purpose. For both the bootstrapping and the randomization method, the percentage of p-values below the unadjusted and the adjusted level of significance, and the median and standard deviation of p-values were determined using the Wilcoxon rank-sum test, to assess the test characteristics of power (sensitivity) and significance level (specificity) for each measure.

RESULTS

The no-JSN participants displayed a marginally lower age (60.6±9.0 vs. 64.2±9.4 years, p=7.5E⁻⁶), body height (166.3±8.7 cm versus 168.2±9.4 cm; p=0.014), and body weight (81.6±15.2 kg vs. 84.4±16.8 kg; p=0.022) than the JSN participants. The difference in BMI (29.4±4.6 kg/m2 vs. 29.8±4.7 kg/m2), however, was not statistically significant (p=0.199). In the subcohort with both MRI and JSW readings (n=290), age was significantly different between knees with baseline JSN versus no-JSN (p=0.013), but there were no significant differences in height (p=0.10), weight (p=0.16), and BMI (p=0.56) between JSN and no-JSN knees.

In the total cohort (n=607), the rate of change in the femorotibial compartments varied from −0.1% in LFTC of no-JSN knees to −1.2% in MFTC in JSN knees (Table 1). The level of statistical significance of the differences in progression between no-JSN and JSN knees was higher for MFTC (p=0.003 without correction for multiple testing) than for LFTC (p=0.090). When analyzing cartilage plates, the rates of change (Table 1) were greater for the femur than for the tibia medially, but were greater for the tibia than for the femur laterally (Table 1). Differences in cartilage thickness loss between JSN versus no-JSN knees were most apparent in the medial femur (cMF; p=0.007 without correction; Table 1).

Table 1. Femorotibial compartments and cartilage plates.

Change in cartilage thickness (ThCtAB) over 12 months in knees without baseline JSN (no-JSN, n=308) and with baseline JSN (JSN, n=299).

	no-JSN			JSN			Between-Group
	MC [µm]	SD [µm]	MC [%]	MC [µm]	SD [µm]	MC [%]	DIFF [µm]	CI [µm]	p-value	SL
MFTC	−12	100	−0.3	−40	128	−1.2	28	10 / 46	.003*	.025
LFTC	−8	81	−0.2	−29	126	−0.8	21	4 / 38	.090	.025
MT	−2	46	−0.1	−10	55	−0.6	8	0 / 16	.027	.013
cMF	−11	76	−0.6	−30	96	−1.8	20	6 / 33	.007*	.013
LT	−10	48	−0.5	−21	70	−1.1	11	1 / 20	.119	.013
cLF	2	58	0.1	−7	84	−0.4	10	−2 / 21	.201	.013

MC = mean change in µm or %, SD = standard deviation, DIFF = mean difference between changes, CI = 95% confidence intervals of differences between changes (lower / upper limit), SL = Significance level after Bonferroni-Dunn correction: The significance (p value) of the differences between changes in JSN and no-JSN knees was computed using non-parametric Wilcoxon rank-sum tests and is reported in the table without adjustment for multiple comparisons. Compartments and cartilage plates showing significant differences after Bonferroni-Dunn correction (p<0.025 for compartments, p<0.0125 for cartilage plates) are marked with *.

MFTC = medial femorotibial compartment, LFTC = lateral femorotbial compartment; MT = medial tibia, cMF = central, weight-bearing part of the medial femoral condyle, LT = lateral tibia, cLF = central, weight-bearing part of the lateral femoral condyle.

When analyzing femorotibial subregions, the greatest mean changes were observed in the central aspect of the weight-bearing femur medially (ccMF) and in the central aspect of the tibia laterally (cLT) (Table 2). cLT was also the subregion to best discriminate the rate of change between no-JSN and JSN knees (p=0.016 without correction) laterally, whereas the level of significance for cartilage thickness loss in JSN versus no-JSN knees in the medial compartment was higher for the external aspect of the weight-bearing femur (ecMF; p=0.008), the external aspect of the tibia (eMT; p=0.009) and the posterior aspect of the tibia (eMT; p=0.022) than for ccMF (p=0.030) (Table 2). In all of the 8 medial, and in 13 of the 16 total (medial and lateral) subregions, the rates of cartilage thickness loss were greater for JSN than for no-JSN knees (Table 3).

Table 2. Femorotibial subregions.

Change in cartilage thickness (ThCtAB) over 12 months in knees without baseline JSN (no-JSN, n=308) and with baseline JSN (JSN, n=299).

	no-JSN			JSN			Between-Group
	MC [µm]	SD [µm]	MC [%]	MC [µm]	SD [µm]	MC [%]	DIFF [µm]	CI [µm]	p-value	SL
cMT	−8	91	−0.3	−24	103	−1.0	16	0 / 31	.042	.003
eMT	−6	80	−0.4	−21	88	−1.6	15	1 / 28	.009	.003
iMT	−4	65	−0.2	−8	70	−0.4	4	−7 / 15	.331	.003
aMT	4	64	0.2	3	69	0.2	1	−10 / 11	.775	.003
pMT	1	57	0.1	−8	61	−0.5	9	0 / 18	.022	.003
ccMF	−23	122	−1.0	−49	150	−2.5	26	4 / 48	.030	.003
ecMF	−4	84	−0.2	−22	105	−1.7	19	4 / 34	.008	.003
icMF	−7	69	−0.4	−22	86	−1.2	15	2 / 27	.031	.003
cLT	−21	98	−0.7	−49	139	−1.8	28	9 / 47	.016	.003
eLT	−6	67	−0.3	−16	81	−1.1	11	−1 / 23	.270	.003
iLT	−16	70	−0.8	−29	100	−1.7	13	−1 / 27	.220	.003
aLT	0	61	0.0	−5	75	−0.3	5	−5 / 16	.235	.003
pLT	−11	91	−0.6	−11	105	−0.6	0	−16 / 16	.715	.003
ccLF	2	85	0.1	−14	127	−0.6	15	−2 / 33	.123	.003
ecLF	4	71	0.3	−6	90	−0.4	10	−3 / 23	.279	.003
icLF	0	69	0.0	−4	85	−0.3	5	−8 / 17	.447	.003

MC = mean change in µm or %, SD = standard deviation, DIFF = mean difference between changes, CI = 95% confidence intervals of differences between changes (lower / upper limit), SL = Significance level after Bonferroni-Dunn correction: The significance (p value) of the differences between changes in JSN and no-JSN knees was computed using non-parametric Wilcoxon rank-sum tests and is reported in the table without adjustment for multiple comparisons. None of the femorotibial subregions showed significant differences after Bonferroni-Dunn correction (p<0.0031). (c|e|i|a|p = central|external|internal|anterior|posterior subregion of the medial tibia (MT) or lateral (LT) tibia. c|e|i = central|external|internal subregion of the central part of the medial weight-bearing femur (cMF) or lateral weight-bearing femur (cLF).

Table 3. Femorotibial orders (ordered value approach).

Change in cartilage thickness (ThCtAB) over 12 months in knees without baseline JSN (no-JSN, n=308) and with baseline JSN (JSN, n=299).

	no-JSN			JSN			Between-Group
	MC [µm]	SD [µm]	MC [%]	MC [µm]	SD [µm]	MC [%]	DIFF [µm]	CI [µm]	p-value	SL
Medial approach:
mOV 1	−98	101	−5.0	−125	125	−7.0	28	9 / 46	3.3E−4*	6.3E−3
mOV 2	−62	80	−3.6	−78	85	−4.8	16	2 / 29	.004*	6.3E−3
mOV 3	−33	50	−1.9	−50	70	−3.0	17	7 / 27	.001*	6.3E−3
mOV 4	−13	45	−0.7	−26	60	−1.6	12	4 / 21	.009	6.3E−3
mOV 5	6	42	0.4	−4	52	−0.2	10	3 / 18	.011	6.3E−3
mOV 6	26	42	1.5	17	51	1.0	9	1 / 16	.016	6.3E−3
mOV 7	48	46	2.7	41	59	2.4	7	−1 / 16	.032	6.3E−3
mOV 8	81	52	4.5	76	65	4.5	5	−4 / 15	.198	6.3E−3
Extended approach:
eOV 1	−136	104	−6.5	−181	144	−9.4	45	25 / 65	5.4E−7*	3.1E−3
eOV 2	−97	80	−5.3	−126	95	−7.3	29	15 / 43	9.6E−7*	3.1E−3
eOV 3	−69	50	−3.7	−95	78	−5.4	27	16 / 37	5.7E−7*	3.1E−3
eOV 4	−53	43	−2.8	−73	61	−4.2	20	12 / 29	1.5E−5*	3.1E−3
eOV 5	−40	40	−2.1	−57	55	−3.2	17	9 / 25	9.5E−5*	3.1E−3
eOV 6	−29	39	−1.5	−43	50	−2.5	15	8 / 22	2.2E−4*	3.1E−3
eOV 7	−19	37	−1.0	−31	45	−1.8	12	6 / 19	.001	3.1E−3
eOV 8	−9	35	−0.5	−19	43	−1.1	10	4 / 16	.006	3.1E−3
eOV 9	1	34	0.0	−7	42	−0.4	8	2 / 14	.032	3.1E−3
eOV 10	10	32	0.5	4	43	0.2	6	0 / 12	.099	3.1E−3
eOV 11	21	32	1.2	16	43	0.9	4	−2 / 10	.340	3.1E−3
eOV 12	31	31	1.7	28	43	1.7	3	−3 / 9	.786	3.1E−3
eOV 13	44	33	2.4	42	44	2.5	2	−4 / 8	.721	3.1E−3
eOV 14	59	37	3.2	61	48	3.5	−1	−8 / 5	.369	3.1E−3
eOV 15	79	43	4.2	81	55	4.7	−1	−9 / 6	.557	3.1E−3
eOV 16	112	55	5.8	116	71	6.5	−4	−14 / 6	.505	3.1E−3

MC = mean change in µm or %, SD = standard deviation, DIFF = mean difference between changes, CI = 95% confidence intervals of differences between changes (lower / upper limit), SL = Significance level after Bonferroni-Dunn correction: The significance (p value) of the differences between changes in JSN and no-JSN knees was computed using non-parametric Wilcoxon rank-sum tests and is reported in the table without adjustment for multiple comparisons. Medial and extended ordered values showing significant differences after Bonferroni-Dunn correction (p<0.0063 for medial, p<0.0031 for extended ordered values) are marked with *.

Ordered values were computed from subregional changes in 8 (medial approach)/ 16 (extended approach) subregions in the medial (medial approach) / medial and lateral (extended approach) femorotibial joint. Ordered value (OV) 1 = subregion showing the most negative change (decrease in ThCtAB) in each subject, OV 2 = subregion showing the second most negative change, ….OV 8 (medial approach) / 16 (extended approach) = subregion showing the smallest negative or the greatest positive change (increase in ThCtAB).

When analyzing medial OVs in the total cohort (n=607, Table 3), 4 showed negative changes (cartilage thinning or loss), and 4 positive changes (cartilage thickening) in the no-JSN knees, whereas 5 showed negative changes and 3 positive changes in the JSN knees. The most significant difference in the rate of cartilage change between no-JSN and JSN knees was observed in OV1 (p=3.29×10⁻⁴). Medial OV 1 through 3 attained smaller p-values than found for any anatomical subregion, cartilage plate, or compartment (Tables 1 to 3).

With the extended (medial and lateral) approach, 8 OVs showed negative changes and 8 positive changes in no-JSN knees, whereas 9 showed negative changes and 7 positive changes in JSN knees (Table 3). The most significant differences in the rate of cartilage change between JSN and no-JSN knees was again observed for OV1 (p=5.38×10⁻⁷). Extended OVs 1 through 7 attained smaller p-values than found for any anatomical subregion, cartilage plate, or compartment (Tables 1 to 3) and OV1 through OV6 for the extended approach displayed a smaller p-value than any OV for the medial approach. The frequency with which the subregions represented extended OV1 was not uniformly distributed and ranged between 2.6% (eMT, pMT & ecMF) and 14.6% (cLT) in the no-JSN sample and between 1.7% (aMT) and 16.1 % (ccMF) in the JSN sample.

When correcting the observed p-values for multiple parallel testing at the compartment (2 compartments) or plate level (4 plates), a significantly different rate of change between JSN and no-JSN knees was observed in MFTC and cMF. In the 16 subregions, none of the changes differed significantly between JSN and no-JSN knees after Bonferroni-Dunn correction. In contrast, 3 of the 8 medial OVs, and 7 of the 16 extended OVs differed significantly between JSN and no-JSN knees. None of the OVs with positive changes (cartilage thickening) displayed significant differences between JSN and no-JSN knees after Bonferroni-Dunn correction, both for the medial and for the extended approach.

In the subcohort of knees with quantitative measurement of the radiographic mJSW (n=290), the rate of change in mJSW did not differ significantly (p=0.386) between JSN and no-JSN knees (Table 4). Medial OV1 (p=5.12×10⁻⁴) and the extended OV1 (p=2.10×10⁻⁵), however, significantly discriminated rates of progression between JSN and no-JSN knees even after Bonferroni-Dunn correction for multiple testing.

Table 4.

Change in minimal medial joint space width (mJSW), medial compartment cartilage thickness, and ordered values of subregional cartilage thickness (ThCtAB) change over 12 months in knees without baseline JSN (n=143) and with baseline JSN (n=147) for which longitudinal mJSW and the MRI outcomes were available.

	no-JSN			JSN			Between-Group
	MC [µm]	SD [µm]	MC [%]	MC [µm]	SD [µm]	MC [%]	DIFF [µm]	CI [µm]	p-value	SL
mJSW	−81	575	−1.7	−129	654	−3.3	48	−94 / 190	.386	.050
MFTC	−18	113	−0.5	−59	142	−1.7	41	11 / 70	.007*	.025
MT	−4	50	−0.2	−12	59	−0.7	8	−5 / 21	.160	.013
cMF	−14	83	−0.7	−47	110	−2.7	32	10 / 55	.002*	.013
mOV 1	−102	107	−5.1	−146	152	−8.2	43	13 / 74	5.1E−4*	6.3E−3
eOV 1	−140	110	−6.4	−198	169	−10.4	58	25 / 91	2.1E−5*	2.9E−3

MC = mean change in µm or %, SD = standard deviation, DIFF = mean difference between changes, CI = 95% confidence intervals of differences between changes (lower / upper limit), SL = Significance level after Bonferroni-Dunn correction: The significance (p value) of the differences between changes in JSN and no-JSN knees was computed using non-parametric Wilcoxon rank-sum tests and is reported in the table without adjustment for multiple comparisons. Parameters showing significant differences after Bonferroni-Dunn correction (p<0.025 for compartments, p<0.0125 for cartilage plates, p<0.0063 for medial, and p<0.0031 for extended ordered values) are marked with *.

Medial ordered value (OV) 1 = subregion showing the most negative change (decrease in ThCtAB) in the medial femorotibial compartment of each subject. Extended OV 1 = subregion showing the most negative change (decrease in ThCtAB) in the medial or the lateral femorotibial compartment of each subject.

The percentage of p-values below the adjusted significance level of 0.05 was higher for the first 5 extended OVs (range 80.1 % to 96.7%) than for any of the other measures (range 0.4% to 77.6%), when comparing changes of JSN versus no-JSN knees using the bootstrapping method. The distribution of p-values varied from unimodal distributions for parameters showing a high percentage of p-values below the significance level (e.g. extended OV 1–3) to an approximate uniform distribution (e.g. aMT, see Table 5 & Figure 3). The median p-value obtained from the within-group randomization (Table 5) was between 0.49 and 0.51 with a standard deviation of 0.29 for all measures. All distributions of p-values approximated the uniform distribution (see Table 5 & Figure 3).

Table 5.

Percentage of p-values less than 0.05 computed using the bootstrapping method between knees without and with joint space narrowing (JSN) and computed using a randomization of knees with JSN.

		Bootstrap (no-JSN vs. JSN knees)				Randomization (JSN knees)
	aSL	% (p<SL)	% (p<aSL)	Median	SD	% (p<SL)	% (p<aSL)	Median	SD
Regional approach:
MFTC	0.0250	84.7	76.9	2.6E-03	0.09	4.9	2.6	0.50	0.29
LFTC	0.0250	34.8	25.4	0.11	0.27	4.9	2.4	0.50	0.29
MT	0.0125	60.4	39.4	0.03	0.19	4.8	1.2	0.50	0.29
cMF	0.0125	77.8	58.6	0.01	0.12	5.0	1.2	0.50	0.29
LT	0.0125	29.8	14.5	0.15	0.28	5.1	1.3	0.49	0.29
cLF	0.0125	22.9	10.6	0.22	0.29	4.7	1.1	0.50	0.29
cMT	0.0031	50.0	16.4	0.05	0.22	5.0	0.3	0.50	0.29
eMT	0.0031	75.0	37.6	0.01	0.13	4.8	0.3	0.50	0.29
iMT	0.0031	16.7	2.5	0.31	0.30	4.6	0.2	0.50	0.29
aMT	0.0031	5.8	0.4	0.49	0.29	4.9	0.3	0.50	0.29
pMT	0.0031	60.1	22.9	0.03	0.18	5.0	0.3	0.50	0.29
ccMF	0.0031	60.7	23.9	0.03	0.19	5.1	0.4	0.50	0.29
ecMF	0.0031	76.7	39.8	0.01	0.12	5.3	0.3	0.49	0.29
icMF	0.0031	59.0	22.3	0.03	0.20	4.9	0.2	0.51	0.29
cLT	0.0031	65.7	28.0	0.02	0.17	4.9	0.4	0.50	0.29
eLT	0.0031	17.7	2.7	0.28	0.30	4.9	0.4	0.50	0.29
iLT	0.0031	19.7	3.1	0.26	0.30	4.4	0.3	0.50	0.29
aLT	0.0031	23.3	4.3	0.21	0.29	5.5	0.4	0.50	0.29
pLT	0.0031	7.6	0.8	0.46	0.30	5.0	0.3	0.50	0.29
ccLF	0.0031	32.3	7.5	0.13	0.27	4.9	0.2	0.50	0.29
ecLF	0.0031	18.9	3.2	0.27	0.30	5.0	0.4	0.50	0.29
icLF	0.0031	10.7	1.2	0.40	0.30	4.6	0.3	0.51	0.29
Medial ordered values:
mOV 1	0.0063	93.7	77.6	4.5E-04	0.05	5.1	0.6	0.50	0.29
mOV 2	0.0063	79.9	52.6	4.9E-03	0.11	5.1	0.6	0.50	0.29
mOV 3	0.0063	89.8	69.8	1.2E-03	0.07	5.3	0.6	0.50	0.29
mOV 4	0.0063	73.7	44.4	0.01	0.14	4.9	0.6	0.50	0.29
mOV 5	0.0063	73.1	43.9	0.01	0.14	4.9	0.7	0.50	0.29
mOV 6	0.0063	70.7	40.6	0.01	0.15	4.7	0.6	0.50	0.29
mOV 7	0.0063	61.9	31.9	0.02	0.18	4.8	0.6	0.49	0.29
mOV 8	0.0063	30.4	10.0	0.15	0.28	5.0	0.7	0.50	0.29
Extended ordered values:
eOV 1	0.0031	99.8	96.7	1.9E-06	0.01	5.0	0.3	0.49	0.29
eOV 2	0.0031	99.7	96.2	2.2E-06	0.01	4.9	0.2	0.49	0.29
eOV 3	0.0031	99.9	97.4	1.0E-06	0.00	5.1	0.3	0.49	0.29
eOV 4	0.0031	98.9	90.0	2.1E-05	0.02	5.0	0.3	0.49	0.29
eOV 5	0.0031	96.7	80.1	1.4E-04	0.03	5.0	0.2	0.50	0.29
eOV 6	0.0031	94.8	74.0	3.1E-04	0.05	5.2	0.3	0.49	0.29
eOV 7	0.0031	90.2	62.2	1.1E-03	0.07	5.1	0.3	0.50	0.29
eOV 8	0.0031	75.3	38.5	0.01	0.13	5.2	0.4	0.50	0.29
eOV 9	0.0031	54.9	19.3	0.03	0.20	4.9	0.3	0.49	0.29
eOV 10	0.0031	36.5	8.8	0.10	0.27	5.1	0.3	0.50	0.29
eOV 11	0.0031	17.0	2.2	0.29	0.30	5.0	0.3	0.50	0.29
eOV 12	0.0031	6.1	0.4	0.47	0.29	5.0	0.3	0.50	0.29
eOV 13	0.0031	7.8	0.7	0.45	0.30	4.7	0.3	0.50	0.29
eOV 14	0.0031	10.8	1.4	0.39	0.30	5.0	0.3	0.51	0.29
eOV 15	0.0031	6.8	0.6	0.46	0.29	4.9	0.3	0.50	0.29
eOV 16	0.0031	7.6	0.7	0.45	0.30	4.8	0.3	0.50	0.29

SL = Significance level (0.0). aSL = Significance level (0.05) adjusted for multiple comparisons. % (p<SL) = Percentage of p-values less than the significance level. %(p<aSL) = Percentage of p-values less than the adjusted significance level (aSL). Median = Median p-value obtained from the p-values computed after each of the 10,000 bootstrapping runs (no-JSN vs JSN knees) and after each of the 10,000 randomization runs (JSN knees vs. JSN knees). SD = standard deviation of the p-values;. All p-values were computed using non- parametric Wilcoxon rank-sum tests. (c|e|i|a|p = central|external|internal|anterior|posterior subregion of the medial tibia (MT) or lateral (LT) tibia. c|e|i = central|external|internal subregion of the central part of the medial weight-bearing femur (cMF) or lateral weightbearing femur (cLF). Ordered values were computed from subregional changes in 8 (medial approach: mOV 1–8)/ 16 (extended approach: eOV 1–16) subregions in the medial (medial approach) / medial and lateral (extended approach) femorotibial joint.

Figure 3.

Graphs showing the distribution of p-values obtained from A) the bootstrapping method and from B) the randomization of changes in JSN knees. The distribution is shown for the entire medial (MFTC) and lateral (LFTC) femorotibial compartment, the anterior subregion of the medial tibia (aMT), the medial ordered value 1 (mOV1), and the extended ordered values 1 and 16 (eOV1, eOV 16).

DISCUSSION

In this study we tested the hypothesis that an extended ordered values approach is superior to conventional approaches of measuring subregional MRI-based cartilage thickness loss, and to radiography, in longitudinally differentiating rates of progression in knees with and without JSN at baseline. Because previous studies have suggested that knees with radiographic JSN at baseline display greater rates of cartilage loss than those without JSN ^7,11,27, this hypothesis was tested in JSN versus no-JSN knees from the OAI. The primary purpose of the study was to explore the gain in sensitivity to differences between groups, when using the extended OV approach to differentiate structural OA progression in two groups with previous evidence of differences in the rate of cartilage loss. It is important to note that this approach is not limited to the question of difference in cartilage loss of JSN versus no JSN knees, but may also be applied to elucidate the impact of other risk factors on OA progression, or to evaluate the effect of a potential DMOAD.

The findings show that the removal of the link between the magnitude of change and its specific location (in any given knee) is highly effective in improving the sensitivity in detecting significant differences in the rates of progression between groups. Particularly in a cohort that includes knees with both medial and lateral radiographic OA (as in the current study), this may be attributed to the fact that changes occur only in some (but not in other) subregions, and that changes across different subregions (in the same or in the contralateral compartment) are not generally positively correlated. Radiography provides a composite measure of cartilage thickness, meniscus integrity and extrusion³¹ and is unable to reveal the spatial heterogeneity of cartilage thickness changes, whereas MRI provides the opportunity to capture change in multiple subregions. However, particular statistical approaches (i.e. OVs) are required, in order to fully exploit the value of subregional information that is provided. A potential downside of the extended (medial and lateral) OV approach proposed is the need to perform segmentations in both femorotibial compartment, which extends analysis time and cost.

In the absence of an external gold standard, the results obtained with the bootstrapping method support the sensitivity levels observed for the different methodologies. The observed percentage of p-values below the defined significance level for the between-group differences was higher for the first five extended OVs than for any other measure. This confirms the higher sensitivity to between-group differences for extended ordered values, even when adjusting for multiple comparisons.

The randomization of the changes within the group of JSN knees showed a similar specificity of all measures used in this study, as the p-values were almost uniformly distributed for all measured parameters and the median and standard deviation observed for each parameter closely approximated the theoretical median (0.5) and standard deviation (0.289) of a uniform distribution between 0 and 1. Because the number of false positives (percentage of p-values smaller than the significance level) was consistent with the defined false positive rate for all of the measures, the medial and the extended OVs can be assumed to display a similar specificity as the regional approach.

A general challenge in designing a clinical trial in OA (either for identifying risk factors or for testing DMOADs) is to determine a primary outcome parameter (i.e. ThCtAB in a compartment, plate or subregion) “a priori”. This is particularly true for MRI, which due to its three-dimensional nature allows for the analysis of changes (either quantitatively or semi-quantitatively) in a multitude of articular tissues, and also in a great number of anatomical subregions ³⁶. Previous studies employing quantitative (sub)regional cartilage analysis with MRI observed that longitudinal changes of cartilage thickness display substantial spatial heterogeneity between knees, and also found variable results between studies in reasonably sized cohorts ³⁷. Another recent study was unable to identify significant change in ThCtAB over relatively short observation periods of 3 and 6 months, despite the fact that the “most progressive” medial subregion (ccMF) was selected as an outcome, and albeit only knees with medial radiographic disease and several risk factors of OA progression were selected ³⁸. To overcome the limited sensitivity to change of quantitative MRI (and radiography) due to spatial heterogeneity of cartilage loss, Buck et al. ¹⁷ proposed an OV approach of subregional changes in MEDIAL compartment cartilage thickness. This approach used only the medial compartment because it was expected to be the region of greatest change in a study of subjects with medial disease, but within the medial compartment change may vary between knees ¹⁷, likely due to the individual mechanical and/or biological conditions. However, it is well known that, in general OA populations, knees preferentially show cartilage loss in the medial or lateral femorotibial compartment, and that limb alignment is the main determinant of medial versus lateral progression ^16,18–21. A recent DMOAD study comparing the sparing effects of licofelone and naproxen in OA defined loss of cartilage volume in the MEDIAL femorotibial compartment as the primary efficacy outcome measure ²². Although this primary outcome measure was reached in this study, the authors reported the protective effect of licofelone to act predominantly in the LATERAL femorotibial compartment. The “extended” OV approach presented here can overcome this challenge in the context of clinical trials ^22,38, as it does not require one to define the primary outcome “a priori” in terms of a specific compartment, cartilage plate or subregion. This not only permits one to widen inclusion criteria during screening for a clinical trial (i.e. to include knees with either medial or lateral disease), but also to generalize the results by allowing one to examine a general OA cohort with few restrictions.

The results from the current study show that the spatial origin of OV 1 is heterogeneously distributed across the joint, but that some subregions are more frequently involved (e.g. ccMF and cLT) than others (i.e. no random distribution). This heterogeneous distribution provides one of the reasons why OV1 is more sensitive to differences (in the rate of change) between groups than region-specific analyzes. The current study therefore shows for the first time that, when using an extended OV approach, the level of sensitivity in differentiating rates of progression between JSN and no-JSN knees is substantially increased over radiography, the analysis of total cartilage plates and compartments, and the analysis of anatomically defined subregions. Additional statistical power may be gained when “a priori” defining OV1 alone as a primary outcome, or when averaging results over a group of orders (i.e. 1 to 4). Averaging changes in cartilage thickness for orders 1 to 4, however, did not provide lower p-values between JSN and no-JSN knees than OV1 alone (data not shown). The results of the current study thus indicate that, in context of baseline radiographic JSN, extended OV1 is the most effective measure in determining differences in rates of cartilage loss. Further studies are required to determine whether this is also true for other risk factors of OA progression, or for treatment with specific DMOADs. If a DMOAD is primarily targeted at reducing cartilage loss, however, OV1 may potentially be used as an effective and powerful “single” outcome measure, whereas definition of a region-based outcome (compartment, plate or subregion) may involve greater needs (and costs) for selecting specific knees, and/or may increase the risk of study failure. Because of its potential for higher statistical power (both with and without correction for multiple testing), the OV approach may become a valuable tool for reducing the number of participants or the observation time in a clinical trial, without unnecessarily sacrificing the generality of the findings.

In conclusion, an extended OV approach, based on medial and lateral femorotibial subregions, showed a higher discriminatory power than radiography, region-based approaches with MRI, and medial OVs when comparing longitudinal cartilage thickness changes in knees with and without baseline JSN in knees from the OA Initiative. Because the (extended) OV approach removes the link between the magnitude and location of change, it allows for the inclusion of knees with both medial and lateral disease, and of knees with different biomechanical risk factors influencing the load distribution within the femorotibial joint. As this circumvents the challenge of selecting a particular knee compartment or anatomical subregion as an outcome measure of progression “a priori”, the approach may generally provide a powerful tool in studies targeting risk factor identification or treatment efficacy in osteoarthritis.