Use of Machine Learning and 71-Plex Immune Mediator Analysis to Identify Nasal Mucus Biomarkers Associated With Olfactory Loss in Patients with CRSwNP

Abstract

Background

The mechanisms driving chronic rhinosinusitis with nasal polyps (CRSwNP)-related olfactory loss remain largely unknown. Here we sought to identify novel modulators of olfactory function via the examination of nasal mucus biomarkers using an expansive 71-cytokine plex analyzed via machine learning models.

Methods

Olfactory testing was performed via 40-question smell identify test (UPSIT). During endoscopic sinus surgery, sponges were placed in the middle meatus of individuals with CRSwNP (n = 15). Nasal mucus samples were screened by multiplex analysis for 71-cytokine/chemokines. Results underwent analysis with statistical and machine learning model approaches to assess whether protein concentrations were predictive of olfactory dysfunction.

Results

In CRSwNP, multiple machine learning models revealed novel cytokines IL-21 and MIP-1δ as positive predictors of greater olfactory dysfunction. Other cytokines detected by more than one model as predictive of olfactory dysfunction were IL-18, MCP-1, IL-22, and BCA-1. Other cytokines identified to be predictive by at least one model were FLT-3L, LIF, IL-20, SCF, IL-23, and TPO.

Conclusion

Using a 71-cytokine/chemokine plex analyzed via machine learning, we identified potentially novel roles for MIP-1δ and IL-21 as modulators of olfactory function in CRSwNP. Use of machine learning for the analysis of nasal mucus cytokines, may serve as powerful tool to analyze complex multiplex immune mediator data.

Introduction

Chronic rhinosinusitis (CRS), a disease affecting 12% of the population, is associated with olfactory dysfunction. ¹ Approximately 88% to 93% of chronic rhinosinusitis with nasal polyps (CRSwNP) patients experience olfactory dysfunction, but the mechanisms that contribute to dysfunction have not been fully elucidated. ² Traditionally, OD in CRSwNP is thought to be due to obstruction of airflow to the olfactory cleft (OC). However, recent research strongly suggests that a combination of etiologies lead to OD, with one of them being localized inflammation occurring at the OC, disrupting mucosa and neuronal transmission.^3,4

Our understanding of CRS pathophysiology has been advanced through the utilization of nasal mucus as a means to study localized inflammation. ⁵ We, and others, have identified numerous immune mediators that change in response to medical or surgical management, and that are predictive of polyp recurrence and are associated with subjective disease severity.^6–8 Additional studies have also identified immune mediators that are associated with OD in CRSwNP.^9,10 Nasal mucus biomarker research often relies on correlations and classical regression models to describe relationships between variables, which are designed to optimize the accounting of in-sample variation. Machine learning (ML) is an alternative approach designed to optimize out-sample predictive accuracy. As we expand our search for potential biomarkers for OD, the benefits of leveraging alternative model families to detect non-linear relationships could be a strong complementary approach.

Previous studies have used a variety of ML models to study OD in the setting of CRS, the most common used to date being the random forests.^11,12 These previous studies either utilize clinical data alone or leverage a small panel of cytokines chosen based on current paradigms of disease. ¹³ In this study, we took “hypothesis free” stepwise regression and ML approaches to study OD in CRSwNP, beginning with a large discovery 71 multiplex cytokine/chemokine/growth factor panel. We hypothesized that this approach would enable us to (a) determine novel candidate biomarkers associated with olfactory loss and (b) suggest next steps to better understand the immunopathology driving CRSwNP associated olfactory loss.

Methods

Study Patient Population

Institutional Review Board granted approval prior to initiation of the study and informed written consent was obtained from all participants. Nasal mucus was collected at the time of endoscopic sinus surgery in patients ≥18 years of age. Inclusion criteria included CRSwNP patients (n = 15) that met the diagnostic criteria outlined by the European Position Paper on Rhinosinusitis and Nasal Polyps 2020. ¹⁴ Exclusion criteria included CRS without nasal polyps, use of immunomodulatory agents within the preceding 30 days, and other immunologic, renal, gastrointestinal, cystic fibrosis, ciliary dyskinesia, endocrine, skeletal disorders, or pregnancy. A total of 15 CRSwNP participants (N = 15) were included in the final analysis. Table 1 contains a full description of demographic factors, comorbidity, and olfactory assessment scores. Regarding comorbidities, asthma was present in 53.3% of the CRSwNP participants.

Table 1.

Demographic and Clinical Characteristics of Participants.

Characteristic	CRSwNP (N = 15)
Age, mean (SD), y	46.3 (18.8)
White	13 (86.7)
Female	6 (40.0)
Smoking (current)	1 (6.7)
Asthma status	8 (53.3)
Diabetes mellitus	2 (13.3)
Allergic Rhinitis	13 (86.7)
Revision Surgery	7 (46.7)
UPSIT	27.1 (8.2)
Normosmia	4 (26.7)
Mild Microsmia	4 (26.7)
Moderate Microsmia	2 (13.3)
Severe Microsmia	2 (13.3)
Total Anosmia	3 (20.0)

Listed at number (percentage) unless otherwise noted. Abbreviations: CRSwNP, chronic rhinosinusitis with nasal polyps; UPSIT, University Pennsylvania Identification Test; SD, standard deviation.

Olfactory Assessment

Objective olfaction was tested after consent, during the pre-operative visit using the University of Pennsylvania Smell Identification Test (UPSIT) (Sensonics International, Haddon Heights, NJ). The UPSIT contains 40 microencapsulated odors released by a scratch-and-sniff method with four answer choices per odor. Olfaction was stratified into the following categories: Normosmia; Mild, moderate, or severe microsmia; and total anosmia. Olfaction diagnosis based on the UPSIT also factored in gender and age percentiles. Patients with scores between 0 and 5 were not included in this study due to probable malingering.

Sample Collection and Analysis

During endoscopic sinus surgery, nasal mucus was obtained by placing polyurethane foam sponges in the middle meatus of study patients bilaterally as previously described.^15–17 After 5 minutes, the sponges were removed then centrifuged at 4 °C for 10 minutes and stored at −80 °C until needed for assays. Samples were shipped to Eve Technologies (Calgary, Canada) who conducted the Human Multiplex Cytokine Array/Chemokine Array 71-Plex Panel (HD71). The panel assessed concentrations of 71 analytes, described in Supplemental Table S1. Each specimen on each panel was run in singlet by Eve Technologies, which is certified to perform high complexity laboratory testing under the Clinical Laboratory Improvement Amendments (CLIA). ¹⁸

Statistical Analysis Software

All analyses were performed in R. ¹⁹ Analysis scripts relied on the Tidyverse collection ²⁰ and additionally on the packages order, ²¹ patchwork, ²² and broom. ²³ ML scripts relied on the Tidymodels collection ²⁴ and additionally on the ordinalNet, ²⁵ rpartScore, ²⁶ and ordinalForest ²⁷ packages.

Pre-Processing

Raw multiplex data were provided by Eve in an Excel spreadsheet expressed as pg/mL. Out of range (OOR) data was then imputed with exact values by Eve. One protein, GROα, yielded entirely out-of-range values that could not be extrapolated. Table 2 contains descriptive statistics for all 71 analytes. All clinical variables, including UPSIT score, were available for all cases.

Table 2.

Comparison of Middle Meatus (MM) Concentrations Between Control and CRSwNP Group (pg/mL).

	CRSwNP (N = 15)
Mediator	Mean	SD	Median	Minimum	Maximum
EGF	872.7	700.6	553.1	220.8	2966
Eotaxin	21.42	8.265	20.33	6.13	31.14
FGF-2	1642	1212	1097	300	4744
FLT-3L	34.56	21.06	31.4	9.34	92.55
Fractalkine	2338	1112	2475	363.5	4706
G-CSF	39188	39468	27552	301.1	138921
GM-CSF	2057	909.7	1840	340.5	3383
GROα	-	-	-	-	-
IFNα2	115.7	56.74	110.1	44.4	215.4
IFNγ	11.29	5.332	10.28	4.02	21.21
IL-1α	13420	7843	11207	4490	29113
IL-1β	299.7	562.7	47.36	6.59	2102
IL-1RA	27846	14146	21737	5733	60016
IL-2	3.332	1.485	3.59	0.56	6.19
IL-3	3.421	1.258	3.21	1.74	6.16
IL-4	11.76	5.912	10.64	1.71	22.81
IL-5	2006	4781	210.7	5.76	17805
IL-6	1666	3429	287.7	27.12	13106
IL-7	56.47	35.63	45.65	4.65	132.4
IL-8	10401	6016	9813	1082	29001
IL-9	57.18	72.58	27.97	8.51	287.9
IL-10	56.37	75.73	25.66	0.14	282.2
IL-12p40	95.65	50.79	95.07	20.23	181.5
IL-12p70	11.47	5.575	10.54	2.85	25.74
IL-13	427.5	575.9	251.7	71.86	2374
IL-15	33.2	19.47	30.5	13.66	82.55
IL-17A	10.87	4.247	12.13	1.15	16.36
IL-17E/IL-25	1487	499.6	1465	644.1	2288
IL-17F	45.79	46.49	22.93	5.22	162.3
IL-18	1867	1166	1979	260.2	3939
IL-22	281.3	134.8	341.4	5.12	439.6
IL-27	538.1	179.7	506.7	306.6	946.1
IP-10	91617	155557	18717	788.2	513447
M-CSF	462.2	339	348.8	152.3	1378
MCP-1	902.1	586.1	1033	242.7	2230
MCP-3	35.11	19.96	29.75	15.77	88.67
MDC	552.6	365.8	533.7	118	1302
MIG/CXCL9	51476	27377	48561	15584	102830
MIP-1α	158.9	164.2	108	51.93	708.4
MIP-1β	311	321.5	243.2	19.29	1150
PDGF-AA	827	779.7	584.7	40.5	2867
PDGF-AB/BB	1167	422.2	1349	403.1	1814
RANTES	227.1	400.8	0.68	0.29	1448
sCD40L	230	108.8	261.8	16.75	358.4
TGFα	293.4	125.2	311.7	89.28	612.9
TNFα	151.3	79.56	128.3	37.42	362.9
TNFβ	39.31	18.96	43.77	13.13	68.65
VEGF-A	3806	2289	3599	1139	7955
6CKine	136.9	79.26	140.7	11.07	269
BCA-1	2145	1776	1347	189.9	4491
CTACK	96.95	71.16	79.7	4.77	277.4
ENA-78	6050	5318	4500	322.1	19924
Eotaxin-2	2555	4348	1306	273.1	17846
Eotaxin-3	5140	10121	450.6	7.66	29635
I-309	3.987	12.03	0.8	0.46	47.44
IL-16	2141	1266	2317	173.8	3973
IL-20	598.1	468	654.5	24.41	1596
IL-21	6.916	1.337	6.62	5.87	11.41
IL-23	1035	2194	124.3	24.41	7489
IL-28A	496.8	1231	2.44	2.44	3545
IL-33	1159	622.6	1235	8.38	2483
LIF	13.59	9.655	12.89	1.73	35.1
MCP-2	45.32	29.18	30.6	6.18	115.2
MCP-4	7.479	10.51	2.66	0.43	39.84
MIP-1δ	5005	5443	2991	942	22644
SCF	22.59	23.34	17.4	1.38	80.49
SDF-1α+β	849.8	569.9	885.5	24.41	1910
TARC	63.88	91.26	34.21	4.97	376.8
TPO	133	313.5	9.38	3.97	950.2
TRAIL	5121	3392	4351	1119	13702
TSLP	6.466	8.762	3.15	2.79	28.8

*All protein concentrations are expressed in picograms per mL (pg/mL).

Exploratory Analysis

We first assessed the protein concentration data for normality. We performed D’Agostino tests and Shapiro–Wilk tests on the 15 values for each protein, as well as on values transformed by the common offset logarithm $f(x)=\log (x+1)$ . The results led us to adopt the transformation throughout. We then used scaled principal components analysis (PCA) to obtain a two-dimensional projection of the 15 cases and 47 protein concentrations with numeric values for all cases. To anticipate the potential for these concentrations to discriminate among the UPSIT categories, we generated a biplot with markers color-coded by category. For readability, we calculated a measure of each analyte's contribution (Where the $k$ th principal component ${\mathbf{w}}_k$ accounts for variance ${\lambda }_k$ , and the $j$ th variable ${\mathbf{v}}_j$ has loading $a_j$ onto ${\mathbf{w}}_k$ , so that ${\mathrm{proj}}_{{\mathbf{w}}_k}({\mathbf{v}}_j)=a_j\frac{{\mathbf{w}}_k}{{\mathbf{w}}_k}$ , we calculate the total contribution of ${\mathbf{v}}_j$ as $\sum _k{a}_j{\lambda }_k$ , taking the sum over all principal components.) and included variable axes only for the 8 top contributors.

Predictive Modeling of Olfactory Dysfunction

We took four predictive modeling approaches to identify determinants of olfactory function. While we report results obtained using ordinal models, we also performed an analogous predictive analysis using binomial models on a dichotomized severity score (moderate or severe hyposmia or total anosmia vs normosmia or mild hyposmia). The results were much weaker (effects less detectable and predictions less accurate), which we attribute to the ability of the ordinal approach to leverage the greater heterogeneity in the five ordered UPSIT categories, as has been well documented with respect to dichotomizing continuous variables.

We first used linear discriminant analysis (LDA) to quantify the loading of each protein concentration profile onto the discriminant axes separating the UPSIT categories. We visualized the LDA in a biplot of the first two axes, formatted like the PCA biplot. Second, we used single-predictor (bivariable) proportional-odds ordinal regression models to regress olfaction on individual proteins. These models rely on the parallel regression assumption that the log-odds of each cumulative probability regress linearly on the predictor(s) along the same coefficients. We performed a Brant test to determine whether the observed deviations from our Ordinal Logistic Regression model are larger than what could be attributed to chance alone..^28,29 Because the severity measure is ordinal, so that no single test of distributions can be performed for all cohorts, these models also serve to test which proteins are differentially expressed by severity of loss. Third, we obtained stepwise ordinal regression models via two-way variable selection based on the Akaike information criterion (AIC). New predictors were considered in order of the deviance they explained in the single-variable models. Finally, we used 12 repetitions of 3-fold cross-validation to tune four families of ordinal predictive models: stepwise ordinal regression, elastic net regression, ordinal classification tree, and ordinal random forest. We evaluated models on four performance metrics—accuracy, area under the receiver operating characteristic curve (AUROC), root mean squared error (RMSE) and mean absolute error (MAE)—and we subsequently optimized model hyperparameters for all of these. We fitted the optimized models to the full data set and calculated importance measures for those predictors included in final models. The models, hyperparameters, performance metrics, and importance measures are summarized in Supplemental Table S2.

To summarize the novelty versus redundancy of the two multivariable ordinal approaches (stepwise regression and ML), we tabulated the overlap in incidence of each predictor in the final models. We have no strong reason to suspect that either approach more successfully identifies analytes that will prove useful in future experiments, so we prioritize analytes identified by both approaches, or by the most models, for detailed discussion and follow-up work.

Correlation Analysis With Subjective and Objective Disease Severity

Statistical analysis was conducted using the GraphPad Prism 10.6 software. A D’Agostino and Pearson omnibus test was used to determine if data sets were normally distrusted. A Spearman's correlation analysis was used to determine if a significant correlation existed between patient-matched mucus immune mediators’ levels versus SNOT22 and Lund–Mackay score.

Results

Exploratory Analysis

While both D’Agostino and Shapiro–Wilk tests yielded evidence against normality for both provided and log-transformed values, evidence was much weaker for the latter (Supplemental Figure S1), so we used log-transformed values in subsequent analyses. PCA did not capture most variation in the data along two principal axes, and the biplot did not reveal any patterns with respect to olfactory dysfunction score (Figure 1).

Figure 1.

Principal component analysis (PCA) row-principal biplot of 71 analytes on all CRSwNP cases. Case markers are color-coded by UPSIT category. Axes are plotted for only the top 8 contributing variables. The two-dimensional projection accounts for 29.8% (PC1) + 16.4% (PC2) = 46.2% of the inertia. Positions revealed no obvious clustering or other patterns by OD.

Prediction of Olfactory Dysfunction With Classical Statistics

LDA failed to separate the cases by UPSIT category, in the sense of nonoverlapping convex hulls. The 8 analytes that contributed most to the discriminant axes are IL-27, IL-21, IL-12p40, IL-12p70, IL-17E/IL-25, FLT-3L, IL-3, and PDGF-AB/BB (Figure 2). Loadings of all analytes onto all 4 discriminant axes are reported in Supplemental Table S3.

Figure 2.

Linear discriminant analysis (LDA) row-principal biplot of 71 analytes on all CRSwNP cases. Case markers are color-coded by UPSIT category with stars linking cases to category centroids. Axes are plotted for only the top 8 contributing variables. The plot reveals no clear associations with OD.

Single-predictor bivariate regression found several cytokines to be predictive of olfactory dysfunction. As shown in Table 3, IL-23, IL-21, LIF, MCP-1, IL-18, FLT-3L, SCF, TPO, and IL-20 predicted more severe categories. Bidirectional stepwise modeling found IL-21, IL-22, MIP-1δ, and BCA-1 predictive of dysfunction in CRSwNP patients (Figure 3 and Supplemental Table S4). Only IL-21 carried over from single-predictor to stepwise regression. We found no evidence that our data violated the parallel regression assumption (Supplemental Figure S2 and Table S5). These conventional methods demonstrate that prediction of olfactory dysfunction is possible, and an effect of these predictions can be placed on individual predictors.

Table 3.

Bivariate Generalized Linear Regression Results for the Control and CRSwNP Groups Using a Single Predictor for Olfactory Dysfunction.

Protein	Coefficient	P-Value
IL-23	0.971	.013
IL-21	27.473	.021
LIF	2.166	.021
MCP-1	2.028	.025
IL-18	-1.803	.025
FLT-3L	-2.225	.031
SCF	0.997	.032
TPO	0.924	.036
IL-20	1.035	.037

Listed in order of increasing significant P-value limited to P < .05. Coefficient determines the strength and direction of association. Includes proteins whose two standard error confidence intervals excluded zero.

Figure 3.

Effect estimates and two standard error intervals for analytes included in the two-way stepwise ordinal regression of UPSIT category. Positive (negative) effects predict more (less) severe OD.

Prediction of Olfactory Dysfunction With Machine Learning

We hypothesized that analytes identified by both classical methods and ML may enhance the predictive and mechanistic value of analytes identified. Therefore, we opted to use a number of ML models, which included ordinal regression, ordinal elastic net, ordinal classification trees, and ordinal random forest models, to compare and identify predictive immune mediators associated with olfactory dysfunction.

Only IL-21 was selected by all approaches, and no other analyte was detected by both in-sample–optimized approaches or by all three out-sample–optimized approaches (Figure 4 and Supplemental Table S6). The ordinal tree also failed to identify any predictors. Five additional analytes were identified by ordinal regression models of both optimization types: IL-18, MCP-1, IL-22, MIP-1δ, and BCA-1. Both IL-22 and MIP-1δ were identified by both an in-sample–optimized GLM and an out-sample–optimized non–GLM, along with six other analytes: FLT-3L, LIF, IL-20, SCF, IL-23, and TPO. Finally, four additional analytes were identified by both GLM and non-GLM optimized out of sample: MIG/CXCL9, EGF, IL-2, and IL-33. Many predictors were involved in regulating or inducing inflammatory responses (IL-21, IL-18, IL-23), chemotaxis (MCP-1, MIP-1δ, CXCL9), or tissue homeostasis (IL-22, SCF, EGF). Data summarizing predictors identified by each model are shown in Supplemental Table S6.

Figure 4.

Variable importance plot for optimized model families fitted to all 15 cases models. Positive (negative) sign indicates the analyte is predictive of greater (lower) OD. Predictors are sorted first by the number of models that identified them, then by which models identified them (priority from left to right), and finally by importance. All models were subjectively tuned optimized for performance metrics.

Lastly, to confirm the relationship with olfactory loss was not being driven solely by more severe disease, we examined the relationship between our top six mediators with SNOT22 and Lund-MacKay score correlations. As shown in Supplemental Table S7, none of these mediators significantly correlated with either subjective or objective disease severity. These results would suggest that the associated with olfactory loss, may be independent of disease severity.

Discussion

The precise inflammatory mechanisms that contribute to olfaction loss are not fully elucidated. ³⁰ Continued discovery of inflammatory mediators that predict olfactory dysfunction is of high value as development of targeted therapy with biologics continues to grow. Here we utilized a 71 multiplex discovery panel as a hypothesis free approach to identify candidate biomarkers and/or novel mechanistic pathways predictive of olfactory dysfunction. To optimize analysis of this large dataset of cytokine/chemokine and growth factor panel we leveraged classical statistics and ML techniques to optimize our discovery approach.

Use of classical statistical and complementary ML approaches allowed further determination of which predictors are mechanistically significant in olfactory dysfunction. While several potential predictors were identified, all three machine-learning models converged on elevated MIP-1δ (CCL15) as being associated with more severe olfactory loss. To our knowledge, this is the first report of alterations in MIP-1δ in patients with CRSwNP and the first to identify an association between MIP-1δ and olfactory dysfunction. MIP-1δ is a member of the CC chemokine family that signals primarily through CCR1 and CCR3 receptors, activating G-protein–coupled cascades that stimulate calcium influx and MAPK pathway signaling (ERK1/2, p38), leading to leukocyte recruitment and expression of adhesion molecules and inflammatory mediators.^31,32 A wide variety of cell types can secrete MIP-1δ, although in the lungs of asthmatic patients the predominant source is thought to be eosinophils, ³³ with additional contributions from smooth muscle cells and basophils. Consistent with this, omalizumab therapy has been shown to reduce MIP-1δ levels along with other type-2 immune mediators, ³⁴ suggesting a potential link between IgE-dependent inflammation and CCL15 signaling.

Although MIP-1δ has not been previously reported in CRSwNP, its established role in eosinophil chemotaxis, macrophage activation, and airway remodeling suggests a plausible mechanistic link to sinonasal inflammation. In the context of olfactory loss, MIP-1δ–driven CCR1 activation may promote eosinophil and macrophage infiltration into the OC, leading to secondary cytokine release, epithelial disruption, and neuronal injury. Together, our novel findings relating MIP-1δ elevations to olfactory loss position MIP-1δ as a novel immunologic and potential neuroinflammatory mediator in CRSwNP and highlight the need for further studies to define its role in mucosal–neuronal crosstalk and its potential utility as a biomarker for disease severity or therapeutic response.

IL-21 was the other novel immune mediator shown to have a predictive association with olfactory loss in all three models. While changes in IL-21 expression have been previously reported in CRSwNP, to our knowledge, this is the first report linking it to olfactory function. In 2015, two reports from Asia demonstrated elevated IL-21 in serum and sinonasal tissues, ³⁵ with serum elevations correlating with more severe objective disease severity. ³⁶ Additionally, mechanistic studies identified elevated CD4 T-cells as the predominant source of IL-21 in sinonasal tissue of CRSwNP patients, ³⁷ consistent with other reports examining the cellular sources of IL-21.^38,39 More recent work has expanded this concept, demonstrating that IL-21 expression in eosinophilic CRSwNP drives pyroptosis and loss of Foxp3 Treg cells via Akt–mTOR–NLRP3–caspase-1 signaling, while IL-21 blockade in vivo reduces eosinophilic infiltration and type-2/type-3 cytokine expression. ⁴⁰ Importantly, circulating T follicular helper (Tfh) and T follicular regulatory (Tfr) cells, major cellular sources and targets of IL-21, are both significantly elevated in patients with CRSwNP compared to healthy controls, with a strong positive correlation between their levels. ⁴⁰ This finding suggests a coordinated dysregulation of the Tfh/Tfr axis that may contribute to persistent mucosal inflammation and IL-21 overproduction.

In a further study focused on CRSsNP patients, IL-21 was the only cytokine to correlate with the SNOT-22 extranasal rhinologic symptoms domain, and reduced IL-21 levels were observed in those undergoing revision compared with primary surgery. ⁴¹ Collectively, our findings, along with prior reports, demonstrate that IL-21 may be a key driver of CRS disease severity and from our studies olfactory loss, potentially acting through combined effects on Tfh/Tfr imbalance and Treg instability. However, further mechanistic studies are needed to better understand its potentially differing roles in CRSsNP and CRSwNP.

Our analysis used a multiplex 71-plex cytokine platform coupled with machine-learning–based feature selection to identify predictors most strongly associated with olfactory loss. This unbiased, data-driven approach was designed to move beyond preselected canonical CRS mediators (eg, IL-4, IL-5, IL-13, IL-33) and instead identify novel or underappreciated pathways potentially linked to sensory dysfunction. As a result, cytokines such as IL-21, MIP-1δ, IL-23, LIF, and FLT3-L emerged as stronger discriminators of olfactory loss severity than traditional type 2 cytokines.

We recognize that several well-established CRS cytokines did not appear among the top-ranked predictors. This likely reflects both (a) the statistical nature of our model, which prioritizes features with the highest predictive contribution to the specific endpoint of olfactory function rather than general CRS inflammation, and (b) biological divergence between mucosal inflammatory pathways driving polyp formation and those influencing neuroepithelial integrity or sensory-neuron signaling. Although recent neuroimmune studies have shown that IL-4/IL-13 signaling can directly sensitize sensory neurons and that IL-4Rα blockade can rapidly improve olfaction,^42,43 those effects may occur through transient or highly localized type-2 signaling that is not captured by single time-point olfactory mucus measurements and that may be relatively uniform across CRSwNP subjects. ⁴⁴ In that setting, ML will naturally elevate analytes with greater between-patient variability (eg, IL-21, MIP-1δ, IL-18, BCA-1) because they better discriminate degrees of smell loss. Therefore, we interpret our findings not as contradictory to prior CRS biology, but as highlighting complementary, potentially neuroimmune cytokine pathways that may modulate smell loss in CRSwNP.

There are important limitations to note within this study. The first is that low sample sizes prohibit optimizing ML algorithms through cross-validation. Given the large cost of this screening process, we utilized a cohort of 15 patients. ML models are best utilized using a large sample size as the models can adequately train, test, and more accurately make predictions. However, a goal of the presented study was to apply ML in a translational setting to instead facilitate a hypothesis generating approach that has the capacity to identify novel new potential biomarkers and analytes for mechanistic dissection that can be further investigated in larger cohort studies or validated in mechanistic murine model systems. Lastly, we did not include patients with CRSsNP in this study, so we may not be able to extrapolate our findings to both subsets of CRS.

Conclusion

Here we demonstrate that a hypothesis free biomarker approach combined with ML identified novel analytes, IL-21 and MIP-1δ, that have not been previously associated with driving olfactory dysfunction in CRSwNP patients. Use of ML models provides a complementary novel tool to identify nasal mucus biomarkers beyond the scope of traditional statistics and further facilitates the use of hypothesis free large analyte panels as screening tools aimed at providing insights into the mechanisms of olfactory dysfunction.