Differential diagnosis of otitis media with effusion using label-free Raman spectroscopy: A pilot study

Abstract

Otitis media with effusion (OME) is an important and common condition affecting hearing in pediatric patients characterized by the presence of fluid in the middle ear space. The fluid is normally described as serous or mucoid based on differences in the fluid viscosity. The differential diagnosis of two OMEs, namely serous and mucoid is of significant clinical value because while the former is self-limiting, surgical procedure is commonly required for the latter. However, accurate identification of fluid types remains a challenging target unattainable with current clinical modalities due to unavailability of nonperturbative molecular tools. Here, we report an emerging spectroscopy approach featuring Raman scattering and multivariate analysis of spectral patterns to discern serous and mucoid fluids, obtained from pediatric patients undergoing myringotomy and tube placement, by providing information of differentially expressed molecules with high specificity. We demonstrate the feasibility of Raman spectroscopy-based approach to categorize middle ear effusion based on the characteristic spectral markers, notably of mucin, with classification accuracy of 91% and 93% for serous and mucoid, respectively. Our findings pave the way for further development of such a tool for fully noninvasive application that will lead to objective and accurate diagnosis thereby reducing unnecessary visits and surgical procedures.

Graphical Abstract

1 |. INTRODUCTION

Otitis media with effusion (OME) is the most common cause of reversible hearing impairment in the pediatric population. Persistent or frequently recurrent OME has been associated with negative impact on speech, language and academic progress. OME is characterized by the presence of noninfected fluid in middle ear space behind an intact tympanic membrane. In contrast to acute otitis media (AOM), where the patient shows signs and symptoms of an acute infectious process, conductive hearing loss is the main symptom in the case of OME that can often be under-diagnosed. OME can be broadly subclassified as serous or mucoid based on the flow properties of the middle ear fluid, with serous being water-like and mucoid present as relatively viscous displaying gel-like characteristics. There are important clinical implications of the viscosity characteristics of middle ear fluid since it is commonly agreed that, unlike serous effusion, mucoid effusion does not resolve on its own and often requires myringotomy and tube placement.

The viscosity of the OME has been primarily attributed to mucin [1, 2]. Mucins, which are highly glycosylated large proteins (10–40 MDa) secreted by epithelial cells, represent the major constituent of the entangled viscoelastic gel-like secretion. Mucins principally serve to provide selective molecular barriers at epithelial surface and play a critical biological role in cell signaling [3]. The over or mis-expression of mucin has been reported to play a key role in several cancer types including pancreatic [4] and breast cancers [5]. Specific to our pathology of interest, the mucin concentration has been found to be significantly higher in mucoid effusions when compared to effusions with serous characteristics. Other constituents, notably lipids, salts, macromolecules, cellular debris and water, combine with mucins to form a nanoscopically heterogeneous environment [6].

White light pneumatic otoscopy examination is the current recommended modality for the diagnosis of OME by the American Academy of Otolaryngology and the American Academy of Pediatrics. However, this technique suffers from significant interobserver variability and poor diagnostic certainty in the hands of primary care providers [7]. Current guidelines recommend the use of tympanometry to assist in the diagnosis of OME when pneumatic otoscopic examination is equivocal [8]. Misdiagnosis of OME and AOM may lead to unnecessary surgical procedures and medical follow-ups on one hand as well as severe complications such as mastoiditis, meningitis or brain abscesses on the other. However, current otoscopic examinations are limited to only confirming the presence or absence of middle ear fluid and cannot provide any information regarding its viscosity or chemical composition. Developing a diagnostic method that can provide objective molecular information would, therefore, considerably bolster the clinician’s ability to diagnose and manage OME.

In this milieu, several investigations, including from our laboratory, have employed photonic approaches including auto fluorescence [9], multicolor reflectance [10], diffuse reflectance spectroscopy [11], optical coherence tomography [12–14] and short-wave infrared imaging (SWIR) [15] to better understand middle ear disease processes. Alternately, Raman spectroscopy, owing to its high molecular information content [16], offers a promising, yet underexplored, tool for diagnosis of such pathologies. We have recently evaluated the potential of Raman spectroscopy to differentiate two morphologically similar middle ear lesions, namely cholesteatoma and myringosclerosis [17].

The current study is undertaken to assess its feasibility in accurately classifying middle ear effusion (MEE) as serous and mucoid. Importantly, we seek to elucidate the differential compositional contributors in the MEE, as determined through multivariate analysis of the Raman spectral profiles. This proof-of-the-principle ex vivo study focuses on analyzing undiluted MEE samples at physiologically relevant conditions that were obtained from pediatric patients undergoing myringotomy and tube placement. Our observations reveal the potential of Raman spectroscopy to provide objective assessment of MEE types, based on the measured content of carbohydrate moieties, paving the way for further investigations, especially in the in vivo domain.

2 |. METHODS AND DATA ANALYSIS

2.1 |. Preparation of commercially procured mucin specimens

Mucin used in this study was procured from Sigma-Aldrich (St. Louis, Missouri) and used as received. For Raman acquisition, 20 mg of mucin was added to 1 mL of double distilled water. The mixture was vortexed at 4°C for 30 minutes.

2.2 |. MUC5B collection and purification

Whole human saliva was collected from two volunteers using gentle vacuum suction from unstimulated submandibular glad secretions under the tongue. MUC5B was purified from collected saliva using a Bio-Rad NGC Fast protein Liquid Chromatography system equipped with a Sepharose CL-2B resin-packed column (GE Healthcare Bio-Sciences, Pittsburgh, PA, USA), according to the methods outlined by Frenkel and Ribbeck [18]. Collection and purification were carried out by Dr. Kathryn B. Dupont in the laboratory of Prof. Katharina Ribbeck (Massachusetts Institute of Technology, Department of Biological Engineering).

2.3 |. MEE samples

The study was performed in accordance with the protocol approved by the Institutional Review Board (#09113, approval date: March 15, 2017) at Connecticut Children’s Medical Center. MEE samples were collected from 27 children in Juhn Tym-Taps after myringotomy. The inclusion of pediatric patients in the study was limited to those undergoing myringotomy under general anesthesia. Each sample was obtained from one ear except one patient who had bilateral effusion in which case effusion samples were collected from both ears. Samples were taken at the time of the surgical procedure to place ear tubes since it is the only time we can incise the tympanic membrane to remove the middle ear fluid. Any MEE sample with visible blood content was excluded from the study. The samples were classified in two groups, namely serous and mucoid, based on their flow properties on inversion in conjunction with clinical opinion from an experienced otolaryngologist (T.A.V.). These classified samples (11 serous and 16 mucoid) were then snap frozen and stored at −20°C. The samples, accumulated over a period of two months, were shipped on dry ice for the spectroscopic study.

2.4 |. Raman spectroscopic measurements

The effusion samples were thawed at room temperature prior to Raman spectroscopy experiments. Spectroscopic acquisitions were carried out using our home-built confocal, inverted Raman microscope described in a previous publication [19]. Briefly, a Ti:Sapphire laser (3900S, Spectra-Physics, Santa Clara, CA, USA) of 785 nm wavelength was used for excitation. The laser light was focused onto the sample through a water immersion objective lens (1.2 NA, ×60, Olympus UPLASPO60XWIR, Center Valley, PA, USA) and the same objective was used to collect scattered light from the sample. The scattered light delivered through 50 μm multimode fiber was dispersed by a spectrograph (Holospec f/1.8i, Kaiser Optical Systems, Ann Arbor, MI, USA) and finally detected by a TE-cooled, deep depletion CCD (PIXIS:100BR eXcelon, Princeton Instruments, Trenton, NJ, USA). The power at the sample was held constant at 70 mW. The effusion samples were placed on a quartz cover slip and the integration time was 10 seconds for each point. Nine spectral data points were coadded from the effusion unless otherwise noted.

2.5 |. Data analysis

The acquired spectra were imported into the MATLAB 2015b (Mathworks, Inc., Natick, MA) environment for further analysis. The spectra were corrected for the presence of cosmic rays prior to spectral analysis. Spectral background correction, owing to the nonanalyte specific signals originating from the fiber probe and quartz cover slips, was also performed. The resultant spectra were normalized to the intensity of the phenylalanine ring-breathing mode peak (ca. 1002 cm⁻¹). For classification of the specimen, the Raman spectra were first subjected to principal component analyses (PCA). Operating without any a priori knowledge of the samples, PCA seeks to project the spectral data onto a set of linearly uncorrelated (orthogonal) directions, for example, principal components (PCs), such that the variance in the original data can be captured using only a few PCs.

Furthermore, support vector machines (SVMs) were used on the PC score inputs to develop a decision algorithm for classifying different types of MEE, similar to the models developed in our earlier work [20]. SVM is a supervised learning model, which is built on structural risk minimization concepts and can efficiently perform nonlinear classification by implicitly mapping the inputs into high-dimensional feature spaces through a kernel. A radial basis function with a Gaussian envelope was chosen as the kernel, and the kernel parameters were optimized based on an automated grid search algorithm. The output of the SVM-derived decision algorithm was validated against the known class labels. The visual representation of the SVM-based class estimations was created by Orange data mining software (Bioinformatics Laboratory, University of Ljubljana, Ljubljana, Slovenia).

3 |. RESULTS AND DISCUSSIONS

Figure 1A shows the flow attributes upon inversion of serous and mucoid samples that underscore the substantive differences in their respective viscoelastic properties. The reduced specific viscosity of mucoid effusion has been reported to be significantly higher than that of the serous effusion, that is, 0.11 ± 0.015 for mucoid and 0.034 ± 0.008 for serous [1]. While such values are significant, the limited quantities of the MEE specimens often preclude direct measurements using a standard rheometer/ viscometer. Furthermore, the utility of such measurements necessitates disruption of the tympanic membrane to obtain the effusion. Hence, there remains a pressing need for a noninvasive method that can assay the molecular composition, which in turn forms the basis of the fundamentally different viscoelastic properties in the two pathologies. As revealed from previous studies, similar mucus composition and mucin glycoforms contribute to similar rheology, characterized by log-linear shear thinning of viscosity [6]. However, standard otoscopy examinations through the intact tympanic membrane can only identify the presence of fluid but not delineate the type of effusion (Figure 1B).

FIGURE 1.

(A) Flow characteristics of serous and mucoid middle ear effusion samples, respectively, showing the stark difference in their viscoelastic properties. (B) White light image of middle ear space of a pediatric patient with intact tympanic membrane, which does not offer any information regarding the nature of middle ear effusion

Based on the reported compositional differences in OMEs, we hypothesized that Raman scattering measurements may not only facilitate objective detection but also provide insights into the disease’s defining biochemistry. The mucin content in mucoid and serous effusions has been previously reported as 236.3 ± 170.0 and 22.3 ± 13.7mg mL ⁻¹, respectively [2]. In particular, we suspected that the content of carbohydrate would be substantially higher in mucoid specimen as such moieties account for nearly 80% [21] of mucin (the principal constituent of OME). This glycoprotein is composed of threonine and serine residues with O-linked glycans such as N-acetylgalactosamine, N-acetylglucosamine, fucose, galactose and sialic acid (Figure 2A C) [22, 23]. Of the various glycoforms, MUC5B has been identified as the major secreted form of mucin in OME [24, 25].

FIGURE 2.

(A) Cartoon representation of a mucin monomer. (B) Representative region of the peptide chain of MUC5B (residues 850–892) obtained through homology modeling (Swiss-model). (C) GalNAc modification (shown in stick representation) of a mucin sequence (shown in line representation) (PDB ID: 2li2) (D) and (E) heat maps corresponding to the Raman scattering intensity of the 1084 cm⁻¹ galactose band acquired from a 100 × 100 μm area in commercially procured mucin sample and in MUC5B sample purified from human saliva

Here, we first sought to record spectra of commercial mucin (Sigma, type II-derived from porcine stomach) in the liquid state. As discussed by Lai et al., barring few exceptions (such as mucus expectorated by patients with cystic fibrosis and that obtained from the cervicovaginal tract), collection of fresh undiluted human mucus remains an outstanding challenge [6]. Hence, mucin procured from commercial vendors that are derived from porcine stomach represent a widely used experimental substitute. Figure 2D shows the heat map corresponding to the intensity of the 1084 cm⁻¹ galactose band acquired from spectroscopically mapping a 100 × 100 μm area of the hydrated sample. The evident nonuniformity of the spectral measurements highlights the heterogeneity of the commercial mucin specimen that may be attributed to the mixture of mucin glycoforms in this lyophilized formulation along with impurities. A more cogent picture emerges when spectroscopic mapping is used on purified MUC5B sample isolated from human saliva (Figure 2E). These observations raise questions about the viability of using commercially procured mucin, without further purification, as a model for Raman scattering measurements.

Notably, such difficulties were absent not only in the purified MUC5B sample but, importantly, in the clinically obtained serous and mucoid OMEs. Mean Raman spectra (±1 SD) of these OMEs recorded from 11 serous and 16 mucoid specimens are shown in Figure 3. The spectra here have been normalized, coaveraged and offset for visualization purposes but displayed without background correction. The averaged spectra are acquired from 9 distinct spatial locations in the field of view to address any potential heterogeneity. The spectra were normalized to 1002 cm⁻¹ peak that is characteristic of the ring-breathing mode of phenylalanine and is not sensitive to protein conformational state. This normalization process permits easier elucidation of the relative glycan to polypeptide content that, in turn, should facilitate the differentiation of the OMEs, consistent with our hypothesis. As can be observed from Figure 3A,B, Raman spectral profiles of mucoid and serous effusions appear grossly similar. A number of the evident peaks are attributable to the vibrational modes of mucin (Table S1, Supporting Information), which typically forms 1% to 3% by weight of mucosal gels [26–28]. However, the differences in mucin content, while encoded in the Raman spectra, need to be elucidated through application of multivariate approaches—as subtle differences in the OME profiles are masked by the presence of other spectral interferents.

FIGURE 3.

Representative Raman spectra acquired from serous (n = 11) and mucoid (n = 16) effusion samples. The solid profile depicts the mean spectrum of each sample group and the shadow represents ± 1SD

To ascertain potential differences of the glycan markers, we plotted the intensity ratios of the key spectral bands (normalized to the phenylalanine mode) for serous and mucoid effusions (Figure 4). Here, we specifically focus on pathological changes reflected in 1098, 1124, 1172 and 1264 cm⁻¹ peaks that can be assigned to modes of N-acetylgalactosamine and N-acetylglucosamine [29, 30]. The first two intensity ratios show vast differences between serous and mucoid with high degree of statistical significance (P < .005). The differences in 1172 and 1264 cm⁻¹ features also exhibit variations between the OMEs, albeit not at the same level of statistical significance. For the 1172 cm⁻¹ peak, this relative reduction in contrast may be due to the contribution of threonine to this Raman feature. Such changes are not observed by plotting ratios of other peak intensities such as those corresponding to lipids or other protein bands. Overall, these differences in the sugar moieties underpin our initial hypothesis about the encoding of mucin content (and hence pathological changes manifested in OMEs) in the Raman spectra.

FIGURE 4.

Relative intensity ratios of carbohydrate: phenylalanine peak for mucoid (blue) and serous (yellow) specimen. Three asterisks indicate P < .005, and 1 asterisk depicts P < .05

While linking individual spectral markers to sample composition is promising, the critical challenge here is to recognize the OME type. Given the locally varying compositions of the effusions, it is important that the chemometric analysis must be robust to stochastic variance and directly related to the pathological context (and not just the molecular compositions). In order to capture the variance in the spectral dataset, PCA was first employed on the Raman spectra. To model the spectra-pathology relationship, the PC scores were used as inputs for SVM-based classifier that permits nonlinear classification by mapping these inputs into high-dimensional feature spaces [6]. Furthermore, details about SVM-derived classification model development are provided in section 2. Figure 5A offers a visual representation of SVM classification based on a leave-one-spectrum-out cross-validation framework. While there are a few inaccuracies for segmentation of individual spectra, clustering of a large fraction of the spectra acquired from serous and mucoid OME specimen is evident from the figure. The classification accuracies of each type of effusion are presented in the confusion matrix (Figure 5B) and the receiver operating characteristics (ROC) plot is illustrated in Figure 5C. The area under the curve for the latter is determined to be 0.93. This level of detection accuracy (ca. 91% for serous and 93% for mucoid) is significant, particularly since current imaging tools do not afford objective and facile differentiation of these two pathologies.

FIGURE 5.

(A) Visual representation of SVM-based classification using leave-one-spectrum-out framework. (B) Confusion matrix describing the performance of the SVM classifier in differentiating serous and mucoid samples. (C) ROC curve for the SVM classifier for the diagnosis of mucoid specimen. The ROC curve of two indistinguishable populations, represented by the dashed line, is included for comparison. The AUC is 0.93, the AUC for a perfect algorithm is 1

We further sought to understand how such individual spectral classification discrepancies are linked to a patient-specific identification. In other words, we wanted to determine how, similar to a clinical situation, these predictions on a per-patient (rather than on a per-spectrum) basis would affect the classification accuracy. In this case, the pathological label is assigned to the sample based on the class of the majority of the spectra recorded from that sample. Reanalysis of the data revealed that there were three misclassifications out of the 27 total samples. For each of these sample misclassifications, the SVM model misdiagnosed all nine of the corresponding spectra. We note that the flow attributes for these three MEE specimens did not unambiguously place them in either the serous or mucoid category. Importantly, since the per-patient diagnosis mimics the clinical circumstances closely, these results reinforce the potential of Raman spectroscopy as an accurate modality for sensing the type of otitis media.

Building on these results, our future studies will employ a customized fiber probe to measure through the intact tympanic membrane to investigate its feasibility of discerning OME types in situ. We envision that the thin translucent tympanic membrane (~0.1 mm) will not significantly impede Raman measurements, particularly in light of the encouraging advances in spatially offset Raman spectroscopy (SORS) [31]. SORS has recently permitted the detection of Raman signals from depths up to a couple of centimeters by modulating the distance between the illumination and collection fibers, and can be exploited to suppress potential interference from the tympanic membrane in diagnosis of OMEs.

4 |. CONCLUSION

In summary, our proof-of-concept study establishes the utility of leveraging label-free vibrational spectroscopic markers for identifying OME and analyzing its biochemistry. Significantly, our measurements require only approximately 5 uL of the sample thereby allowing testing of clinically obtained specimen without any dilution which has remained a major challenge in utilizing viscosity measurements for routine diagnostic purposes. Overall, our findings reveal the promise of using label-free Raman spectroscopy to obtain compositional information in MEE that can be harnessed to differentiate between serous and mucoid cases in a noninvasive manner.

Supplementary Material

Supplemental Authors Bio

Table S1. Prominent peaks observed in the Raman spectra and their vibrational mode assignment.