2.1.

Clinical Protocol

All aspects of the clinical protocol were reviewed and approved by the University of Michigan Medical School Institutional Review Board (IRBMED). A total of 40 patients with clinical symptoms of chronic knee pain, and who were also scheduled for elective surgical treatment, were enrolled in the study. Preoperative conventional postero–antero radiographs of the symptomatic knee were reviewed by a musculoskeletal radiologist. Radiographs were read using the Osteoarthritis Research Society International atlas for identification of definite osteophytes,⁴⁶ and OA was defined as the presence of a definite osteophyte in any of three views in accordance with American College of Rheumatology criteria for knee OA.⁴⁷ Overall radiographic severity of OA was assessed using Kellgren/Lawrence (K/L) criteria on the postero–antero view only.⁴⁸ Patients were classified into two groups by K/L score. Even though all of the patients presented clinical evidence of OA (knee effusion and pain), x rays of the affected knee joint had found that 17 patients had a low K/L score ( $\mathrm{K}∕\mathrm{L}=0$ to 1) and were placed in the negative for radiological OA (−/ROA) group. Patients with a K/L score above 2 ( $\mathrm{K}∕\mathrm{L}=2$ to 4) were placed in the positive for radiological OA (+/ROA) group. These grouping are congruent with other arthritis studies that use a K/L score.^{12, 15, 49, 50}

Twenty-two patients in the study underwent total knee arthroplasty, 1 patient underwent hemiarthroplasty and 17 patients underwent therapeutic arthroscopy. Patient demographics are shown in Table 1 . At the time of surgery, SF was obtained from the knee by needle aspiration using a size 18 gauge needle and immediately placed into glass tubes containing anticoagulants and protease inhibitors (SCAT-1, Haematologic Technologies, Inc., Essex Junction, Vermont). Additives in SCAT-1 vials were examined using Raman spectroscopy to verify that their Raman spectrum does not overlap with SF Raman bands. Macroscopic debris in SF was removed by low-speed centrifugation, and specimens were aliquoted and then snap frozen and stored at $-80°\mathrm{C}$ until use.

Table 1

Demographics of patients in the drop deposition/Raman spectroscopy study. Patients were recruited into the study after presenting clinical evidence of knee osteoarthritis. After informed consent, x rays and synovial fluid aspirates were collected from the afflicted knee.

2.2.

Power Analysis

A total of $N=40$ subjects and a normal distribution of K/L scores with 20 subjects in the −/ROA ( $\mathrm{K}∕\mathrm{L}=0$ to 1) group and 20 subjects in the+/ROA ( $\mathrm{K}∕\mathrm{L}=2$ to 4) group was assumed for a priori power calculations. A two-sample $t$ -test of means with a 95% confidence limit yielded a power of 87% $(N=40)$ to detect moderate differences between −/ROA and +/ROA groups. Raman data from three SF specimens were removed from the −/ROA group when comparing Raman band intensity ratios because these specimens had severe contamination from hemoglobin or anticoagulant from the storage vial. Raman data of synovial fluid from 37 patients were used in comparing Raman band intensity ratios between −/ROA $(n=14)$ and +/ROA $(n=23)$ groups. For $N=37$ , the power was 84%. The power for a comparison of all five K/L groups (0 to 4) was 62% using an ANOVA, one-way, fixed effect $f$ -test. The a priori power tests indicated adequate power for comparison of differences between −/ROA and +/ROA groups.

2.3.

Drop Deposition of Synovial Fluid

Synovial fluid from all 40 patients was examined using the DDRS protocol. SF specimens were examined without extensive preparation. Two small-volume $\left(2\text{to}10\mu \mathrm{l}\right)$ drops of synovial fluid were deposited onto a fused silica slide using an Eagle pipette (World Precision Instruments, Sarasota, Florida). Drops were allowed to dry overnight, semicovered, at room temperature and then were examined the following day using light microscopy and Raman spectroscopy. Light microscope images were taken for both SF drops, and Raman spectra were collected from one SF drop.

2.4.

Microscopy/Raman Spectroscopy

A Nikon E600 epi-fluorescence microscope (Nikon, Inc., Melville, New York) was modified for NIR Raman spectroscopy in house. Microscope images were collected in epi-illumination and transmission illumination modes with $2\times ∕0.06\mathrm{NA}$ , $4\times ∕0.20\mathrm{NA}$ , $10\times ∕0.50\mathrm{NA}$ , and $20\times ∕0.75\mathrm{NA}$ S Fluor objectives (Nikon, Inc.). A $785\text{-}\mathrm{nm}$ Kaiser Invictus laser was line focused (Kaiser Optical Systems, Inc., Ann Arbor, Michigan) onto dried drops using a $20\times ∕0.75\mathrm{NA}$ S Fluor objective. Raman-scattered light was collected through the same $20\times ∕0.75\mathrm{NA}$ S Fluor objective and dispersed through a spectrograph (HoloSpec $f∕1.8$ , Kaiser Optical Systems, Inc.). Raman signal was collected for $10\min$ on a charge-coupled device (CCD) detector optimized for NIR wavelengths (DU401-BR-DD, Andor Technologies, Belfast, Ireland). Raman transects consisted of 126 Raman spectra arranged at equidistant points along a line through the specimen. Six to twelve transects were collected at various locations across the surface of each of the dried drops. A total of 418 Raman transects (52,668 spectra) were collected in this study.

2.5.

Data Analysis

Raman data was collected and processed without prior knowledge of the radiography results. Data were imported into MATLAB software (v. 7.0, The Math Works, Natick, Massachusetts) and corrected for curvature, dark current, and variations in the CCD quantum efficiency. A mean spectrum was calculated from each transect after correction for the fused silica background. The mean spectrum was imported into GRAMS/AI (ThermoGalactic, Salem, New Hampshire) and baseline corrected using a user-defined multipoint baseline fitting routine. Baseline-corrected spectra were intensity-normalized to the phenylalanine ring breathing band intensity at $\sim 1001{\mathrm{cm}}^{-1}$ because it was the best resolved band and is not sensitive to local chemical environments. Raman bands in baseline-corrected spectra were fitted with mixed Gaussian/Lorentzian functions. Curve fitting results were accepted when the residuals were minimized $(R^2>0.99)$ and no negative bands were generated. Bands generated in the $600\text{to}1750{\mathrm{cm}}^{-1}$ spectral region were identified as primarily from proteins. Band positions were reproducible to $\pm 2{\mathrm{cm}}^{-1}$ . Raman band data (area, intensity, and width) were imported to Microsoft Excel, and band intensity ratios were calculated. Intensity band ratios at $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ , $1080{\mathrm{cm}}^{-1}∕1125{\mathrm{cm}}^{-1}$ , $1235{\mathrm{cm}}^{-1}∕1260{\mathrm{cm}}^{-1}$ (amide III), $1655{\mathrm{cm}}^{-1}∕1448{\mathrm{cm}}^{-1}$ , and $1670{\mathrm{cm}}^{-1}∕1655{\mathrm{cm}}^{-1}$ (amide I) were used to evaluate protein content and structure. Raman spectra collected from drop edges were analyzed separately from Raman spectra collected from the drop center. After Raman data were analyzed for the entire patient cohort, radiograph scores were provided. Band intensity ratios from 37 patients were plotted according to the patient’s K/L score and compared in two ways: −/ROA group versus +/ROA group, and according to K/L scores 0 to 4. Raman data corresponding to patients with a K/L score of 3 and 4 were grouped together because there were only 2 patients in the $\mathrm{K}∕\mathrm{L}=4$ group. A two-tailed Student’s $t$ -test with unequal variance was used to test for significance in differences between −/ROA and +/ROA groups. Differences in Raman data corresponding with the individual K/L scores (0 to 4) were tested for significance using multiway ANOVA for the major factor (K/L score), and multiple comparison was performed using the “least-significant differences” criterion to examine all possible $t$ -tests between means. Differences were considered statistically significant if the $p$ value was $<0.05$ .

In another test, MATLAB was used to preprocess all spectra from drop edges, perform baseline and background corrections, and generate a mean spectrum for each patient using only the spectra collected from SF drop edges. The mean spectra were input into an unsupervised k-means cluster analysis. Raman peaks were not fitted with Gaussian/Lorentzian functions for k-means clustering. Pixel intensities corresponded to the $1080{\mathrm{cm}}^{-1}∕1002{\mathrm{cm}}^{-1}$ and amide I ratios, and we input two groups to correspond to the −/ROA and +/ROA groups.

3.1.

Light Microscopy of Dried SF Drops

All of the SF drops dried as a heterogeneous deposit. Drop deposition provided a coarse separation, as shown by the low-magnification microscope images in Fig. 1 . In Fig. 1a, the dried drop contains two distinct regions. The outer edge gave a glassy appearance and the drop center contained fern-shaped crystalline deposits. There are other distinguishable features, including cracks radiating from the drop center and interference patterns along the edges of some radial cracks. Most of the drops dried in the same general pattern and were independent of the SF volume aspirated from the patient and K/L score. We did observe some variety in the SF deposition pattern, as seen in Fig. 1. Approximately $1∕3$ of the drops $(n=11)$ did not dry in the deposition pattern seen in Fig. 1a. Alternative deposition patterns either lacked cracks near the drop edges [Fig. 1b], were glassy throughout the drop [Fig. 1c], or contained no fern-shaped crystal deposits [Fig. 1d]. Analysis of microscope images for presence of radial cracks at the drop edge and fern-shaped crystals indicated a moderate correlation between OA severity and the presence of these drop features at low magnification $(R^2=0.30)$ . Detailed visual features are apparent upon higher magnification $(10\times )$ and these features may correlate more strongly with fluid viscosity or arthritis damage. Automated image analysis algorithms, similar to ones used for automated cell phenotype scoring, may also be able to find a stronger correlative trend between the visual features observed in SF drops and radiographic scoring.⁵¹

Fig. 1

Microscope images of human synovial fluid (SF) dried drops at low magnification show a heterogeneous deposit. There were two main regions in the SF drop (a): a smooth glassy film at the drop edge and a fern-shaped crystalline deposit in the drop center. Most of the SF specimens dried in patterns similar to Fig. f1a. There were variations in the drop appearance that included a lack of cracks at the drop edge (b), a glassy film deposit throughout the drop (c), and no fern-shaped crystals in the drop center (d). These variations in drop appearance did not correlate with radiological scoring. When examining drop appearance for fern-shaped crystals or radial cracks, we observed a moderate correlation between the presence of these features and classification into the −/ROA or +/ROA group $(R^2=0.30)$

3.2.

Drop Deposition/Raman Spectroscopy of Dried SF Drops

Raman spectra of the dried SF drops also identified two primary regions within the drop, which supported our findings from the light microscopy studies. Spectra collected from the drop edges were comprised primarily of protein Raman bands, which provided sufficient data for evaluating the physiochemical composition of human SF. Although Raman bands from glycosaminoglycans (GAGs) and lipids can overlap with some protein bands, they are relatively weak Raman scatterers and thus do not contribute significantly to the overall SF spectra. However, interaction of proteins with GAGs and lipids may still influence protein spectral features such as band position, width, height, or area. Previous studies of DDRS to examine solutions of proteins or biofluids found that drop deposition spectra of proteins were similar to solution-state or solid-state spectra.^{41, 43, 44} As expected, SF volume increased with OA damage, but the differences in SF volume were not statistically significant.

To interpret and eventually quantify changes to SF composition, bands were identified, as shown in Fig. 2, and assigned. Table 2 shows band assignments made for SF spectra, based on previous Raman studies of proteins and GAGs.^{52, 53, 54, 55, 56} Figures 3a and 3b show the most promising ratios $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ and amide I band intensity ratios, respectively. The differences between −/ROA and +/ROA groups are significant $(p<0.01)$ . Figures 3c and 3d show the $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ and amide I band intensity ratio sorted by K/L score. The $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ [Fig. 3a] intensity ratio increased in the +/ROA group and is evidence that the chemical environment of the protein backbone is altered with OA damage. The ratio of band intensities in the amide I envelope provide a marker of the relative abundance of disordered random coil secondary structure and ordered $\alpha$ -helix secondary structure. Higher amide I ratio observed in SF from patients in the +/ROA group indicate more disorder in protein secondary structure and provide further evidence of altered electrostatic interactions. Panels (c) and (d) of Fig. 3 show that the mean ratio values begin to overlap in adjacent K/L score groups, but there is a moderate correlation of the ratios with K/L score ( $R^2=0.31$ for amide I, and $R^2=0.35$ for $1080{\mathrm{cm}}^{-1}∕1002{\mathrm{cm}}^{-1}$ ratio).

Fig. 2

Raman spectra collected from SF drop edges contain bands that convey chemical and structural information in the $800\text{to}1750{\mathrm{cm}}^{-1}$ region. Spectra collected from SF were primarily comprised of signal relating to SF proteins. Three sections of the Raman spectra were used to evaluate SF chemistry. Protein backbone information is contained in the $1010\text{to}1150{\mathrm{cm}}^{-1}$ region. The amide envelopes at $1210\text{to}1300{\mathrm{cm}}^{-1}$ (amide III) and $1600\text{to}1710{\mathrm{cm}}^{-1}$ (amide I) relate secondary structure information about protein amide linkages. These regions were found to provide the most distinction between the −/ROA and +/ROA groups.

Fig. 3

Comparison of two band intensity ratios for −/ROA ( $\mathrm{K}∕\mathrm{L}=0$ to 1) +/ROA ( $\mathrm{K}∕\mathrm{L}=2$ to 4) groups found significant differences $(p<0.01)$ in the +/ROA group. Error bars show the 95% confidence interval. The $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ (a) and amide I (b) band intensity ratios show that the C-C and C-N groups in the protein backbone are altered in SF from damaged knee joints, which leads to an alteration in the relative abundance of $\alpha$ -helical versus random coil content. When these ratios are sorted by K/L score [ $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ , Fig. 3c, and amide I, Fig. 3d], these differences begin to diminish, but moderate correlative trends (0.31 and 0.35, respectively) were still observed.

Table 2

Raman band assignments of SF drops were based on literature reports of proteins and biological specimens (Ref. 53, 54, 55, 56). Raman bands in the 1010to1150cm−1 region in the SF spectra were used to evaluate the protein backbone. The protein secondary structure was evaluated using bands in the amide III (1210to1300cm−1) and amide I (1600to1710cm−1) envelopes. Deconvoluted bands in the amide envelopes were assigned to a random coil/disordered protein structure at 1235cm−1 and 1670cm−1 or to an α -helix/ordered protein structure at 1260cm−1 and 1655cm−1 .

Preliminary application of automated data analysis techniques shows a good separation of data in the −/ROA and +/ROA groups. Figure 4 shows k-means clustering of pixel intensities corresponding to the $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ and amide I ratios. The data indicate that simple pixel intensity ratios could be a good input for a k-means cluster analysis. Table 3 presents the sensitivity and selectivity of the cluster analysis. The sensitivity (74%) and selectivity (71%) of the analysis is satisfactory for an unsupervised classification. Using a support vector machine or principle components analysis with cluster analysis, or leave-one-out validation, are possible methods to improve classification rate.

Fig. 4

In another analysis, unsupervised k-means cluster analysis was performed on data collected from the edges of dried SF drops. Pixel intensities corresponding to the $1080{\mathrm{cm}}^{-1}∕1001{\mathrm{cm}}^{-1}$ and amide I ratios and a priori knowledge of two groups were input into the cluster. The two groups were clearly separated; the cluster in the lower-left corner is the −/ROA group, and the cluster in the upper-right corner is the +/ROA group. Classification by k-means cluster was compared against clinical radiological classification. Solid circles represent correct classification, triangles represent false negatives, and squares represent false positives. Sensitivity and selectivity values are provided in Table 3.

Table 3

Results from unsupervised k -means cluster analysis of mean Raman spectra generated from each patient’s synovial fluid (n=37) . Input into the cluster analysis was the ratio of pixel intensities corresponding to the 1080cm−1∕1002cm−1 and 1670cm−1∕1655cm−1 bands and a priori input of two groups. Rows indicate clinical diagnosis; columns indicate predicted class. The index of agreement was 0.46.