Identification of clinical-biological endophenotypes in patients with eosinophilic granulomatosis with polyangiitis using unsupervised machine learning: a real-world study

Abstract

Background

Eosinophilic granulomatosis with polyangiitis (EGPA) is a systemic vasculitis characterized by high clinical heterogeneity. The conventional binary classification based on antineutrophil cytoplasmic antibody (ANCA) status inadequately captures this heterogeneity and provides limited guidance for individualised treatment decisions. This study aimed to identify and validate clinical-biological endophenotypes in EGPA using unsupervised machine learning methods based on multidimensional real-world clinical data, to inform precision medicine approaches.

Methods

In this retrospective single-centre observational study, we included 205 patients diagnosed with EGPA between January 2015 and December 2023 at China-Japan Friendship Hospital. Comprehensive data on demographics, clinical manifestations, laboratory tests, imaging features, and treatment information were systematically collected. Consensus clustering combined with K-means algorithm was applied to identify patient subgroups based on 14 core features, including peripheral blood eosinophil count, fractional exhaled nitric oxide (FeNO), serum total IgE, C-reactive protein (CRP), ANCA status, and multi-system involvement. Principal component analysis was used for dimensionality reduction and visualisation. Differences in clinical characteristics, laboratory parameters, imaging patterns, and treatment responses were compared among the subgroups. An online prediction tool was developed and validated based on the clustering results.

Results

Consensus clustering analysis identified three stable endophenotypes (optimal cluster number k = 3). Subgroup 1 (n = 78, 38.0%): Eosinophilic-Airway Inflammation Type, characterized by markedly elevated eosinophil count (6.10 × 10⁹/L, IQR: 5.10-8.00), FeNO (79 ppb, IQR: 55–112) and serum total IgE (798 IU/mL, IQR: 602–980), predominant respiratory symptoms, with ground-glass opacities and consolidation as main findings on chest high-resolution computed tomography, and among patients who received mepolizumab, a partial remission rate of 89.7% was observed. Subgroup 2 (n = 65, 31.7%): ANCA-Systemic Vasculitis Type, defined by a significantly higher ANCA positivity rate (26.2%) compared to the other subgroups, significant systemic inflammatory response (CRP: 33.5 mg/L, IQR: 22.5–48.5) and frequent peripheral nervous system involvement (76.9%), receiving higher initial glucocorticoid doses. Subgroup 3 (n = 62, 30.2%): Paucieosinophilic-Multi-Organ Damage Type, presenting with relatively low eosinophil-related indices but higher proportions of skin, gastrointestinal, and cardiac involvement, commonly showing fibrotic changes on imaging, and with a numerically lower observed overall response rate to available biologics (51.6%). The developed online prediction tool showed good performance in an independent validation cohort (overall accuracy 90.0%, Kappa agreement with expert judgement 0.850).

Conclusion

Using unsupervised machine learning on real-world data, this study successfully identified three distinct clinical-biological endophenotypes in EGPA, revealing different underlying pathophysiological mechanisms and treatment response patterns, and developed a clinically translatable prediction tool. This endophenotype framework and tool may help inform individualised treatment strategies and provide an exploratory data-driven classification basis and decision support for future precision medicine research in EGPA.

Introduction

Eosinophilic granulomatosis with polyangiitis (EGPA), formerly known as Churg-Strauss syndrome, is a rare systemic vasculitis characterized by eosinophilia, necrotising granulomatous vasculitis, and multi-system involvement [1, 2]. As a member of the ANCA-associated vasculitis (AAV) spectrum, EGPA exhibits a broad clinical phenotype, potentially affecting the respiratory system (e.g., asthma, eosinophilic pneumonia), skin, peripheral and central nervous systems, gastrointestinal tract, heart, and kidneys [3, 4]. This extensive heterogeneity poses significant challenges for early diagnosis, treatment selection, and prognosis assessment.

Current clinical management of EGPA largely relies on the conventional binary classification based on ANCA status [5, 6]. It is generally accepted that ANCA-positive patients are more likely to exhibit features of classic systemic small-vessel vasculitis, whereas ANCA-negative patients more commonly present with prominent eosinophilia and tissue infiltration [7, 8]. However, the clinical utility of this dichotomous approach is increasingly recognised as limited. Numerous clinical observations indicate substantial variations in disease progression trajectories, organ damage patterns, and responses to different therapeutic modalities, particularly to novel biologics such as mepolizumab, even among patients within the same ANCA status group [9–11]. This strongly suggests the existence of distinct patient subgroups within the EGPA disease entity, driven by different immuno-inflammatory pathways and possessing unique clinical-biological features, i.e., “endophenotypes” [12].

Identifying clinical-biological endophenotypes in EGPA is crucial for advancing precision medicine. A robust endophenotype classification system could: (i) more accurately predict the natural history, organ damage risk, and prognosis for different patient subgroups; (ii) guide individualised treatment selection, avoiding the pitfalls of a “one-size-fits-all” approach leading to either under-treatment or overtreatment; and (iii) provide clearly defined target populations and rationales for clinical trial design in the development of novel drugs targeting specific pathophysiological pathways [13, 14]. Previous attempts to subtype EGPA have primarily relied on single dimensions (e.g., ANCA status, eosinophil level, organ involvement pattern) or were based on small observational studies, lacking systematic, multidimensional, and data-driven endophenotype identification [15, 16].

Machine learning methodologies, particularly unsupervised clustering analysis, have emerged as powerful tools for discovering novel endophenotypes in complex heterogeneous diseases such as asthma [17], chronic obstructive pulmonary disease (COPD) [18], and systemic lupus erythematosus (SLE) [19]. These methods can integrate multidimensional data, clinical, laboratory, imaging, to objectively identify stable patient subgroups with shared characteristics from real-world information, without relying on pre-specified hypotheses.

Despite the well-recognised clinical heterogeneity of EGPA, studies exploring its endophenotypes using advanced machine learning techniques based on systematic, multidimensional clinical data remain scarce. Current understanding is largely derived from small retrospective analyses or observations of single biomarkers, lacking comprehensiveness and a data-driven approach [20, 21].

Therefore, this study aimed to address this gap. Utilising a detailed multidimensional clinical database of over 200 EGPA patients accumulated over eight consecutive years at China-Japan Friendship Hospital, we applied robust unsupervised machine learning algorithms, including consensus clustering, for systematic endophenotype exploration. We hypothesised that EGPA patients could be objectively categorised into several endophenotypes with distinct biological significance and clinical utility, based on their clinical manifestations, laboratory inflammatory marker profiles, imaging characteristics, and treatment response patterns. By elucidating these endophenotypes, this study aims to establish the foundation for a more refined and clinically actionable EGPA classification framework, ultimately facilitating a shift towards individualised and precision management of this disease.

Methods

Study design and data source

This was a retrospective, single-centre, observational study. The study population comprised consecutive patients diagnosed with EGPA who attended the Department of Pulmonary and Critical Care Medicine at China-Japan Friendship Hospital between January 1, 2015, and December 31, 2023. The department operates both dedicated outpatient clinics for complex airway and eosinophilic lung diseases and an inpatient ward. Patients were referred or self-presented predominantly due to one or more of the following: (1) abnormal chest high-resolution computed tomography (HRCT) findings such as ground-glass opacities or migratory consolidation; (2) unexplained peripheral blood eosinophilia (typically > 1.5 × 10⁹/L); (3) markedly elevated serum total IgE; (4) deteriorating lung function of unclear etiologies; or (5) difficult-to-control asthma subsequently found to have systemic features. The cohort therefore spans a spectrum of disease severity, from stable outpatients undergoing diagnostic evaluation to patients hospitalized for acute clinical worsening. The diagnosis of EGPA strictly adhered to the 2012 revised American College of Rheumatology classification criteria [22] and considered the 2022 European Alliance of Associations for Rheumatology recommendations [23]. The study protocol was approved by the hospital’s Institutional Review Board, and the requirement for informed consent was waived due to the retrospective nature of the study.

Patient inclusion and exclusion criteria

Inclusion criteria

(1) Age ≥ 18 years; (2) Fulfilling the aforementioned EGPA diagnostic criteria; (3) Availability of complete baseline clinical assessment, laboratory investigations, and chest imaging data; (4) Regular follow-up for ≥ 6 months with documented treatment response evaluation.

Exclusion criteria

(1) Coexisting other definite systemic autoimmune diseases; (2) Active malignancy; (3) Severe organ dysfunction; (4) Pregnant or lactating women; (5) Missing data exceeding 20% for any of the 14 core clustering variables.

A total of 205 eligible EGPA patients were finally included.

Data collection and variable definition

Data were retrospectively extracted from the hospital’s unified electronic medical record system, covering the following dimensions:

Demographics and clinical features

Sex, age at diagnosis, personal and family allergy history. Detailed documentation of organ system involvement, including respiratory (asthma, allergic rhinitis, chronic sinusitis), skin/mucosa, peripheral nervous system, central nervous system, gastrointestinal, cardiovascular, and renal systems. The Birmingham Vasculitis Activity Score (version 3, BVAS-3) was used to quantify systemic disease activity at baseline [24].

Assessment of Cardiac and Peripheral Nervous System Involvement: In our hospital, cardiac involvement was evaluated through cardiology consultation and defined by the presence of one or more of the following attributable to EGPA, after exclusion of coronary artery disease and other alternative etiologies: (a) wall motion abnormalities, pericardial effusion, or valvular regurgitation on transthoracic echocardiography; (b) features consistent with myocarditis on cardiac magnetic resonance imaging; and/or (c) elevated cardiac troponin I (> 99th percentile upper reference limit) in the absence of alternative explanations. For patients without these examinations, cardiac involvement was determined by the treating cardiologist based on clinical assessment. Peripheral nervous system involvement was diagnosed by a neurologist based on characteristic clinical symptoms and signs (most commonly mononeuritis multiplex, presenting as asymmetric motor and sensory deficits in the distribution of individual peripheral nerves) and confirmed by nerve conduction studies and/or electromyography demonstrating axonal neuropathy. Simple paresthesia without objective neurological deficit or abnormal nerve conduction studies was not classified as peripheral nervous system involvement.

Core laboratory indicators

Laboratory results at diagnosis or before treatment initiation. (1) Eosinophil-related indices: absolute peripheral blood eosinophil count (EOS, ×10⁹/L) and percentage (%). (2) Airway inflammation marker: fractional exhaled nitric oxide (FeNO, ppb) as a quantitative measure of type 2 airway inflammation. (3) Systemic inflammation markers: C-reactive protein (CRP, mg/L), erythrocyte sedimentation rate (ESR, mm/h). (4) Autoantibodies: ANCA status. Indirect immunofluorescence was used for initial screening; positive samples were further tested by enzyme-linked immunosorbent assay for myeloperoxidase-ANCA and proteinase 3-ANCA antibody titres. Positivity was defined as a typical IIF pattern with a confirmed ELISA titre ≥ 1:20. (5) Immunoglobulin: serum total immunoglobulin E (IgE, IU/mL).

Imaging features

Baseline high-resolution computed tomography images were independently reviewed by two radiologists with over 10 years of experience in chest imaging, blinded to clinical information. Imaging findings were categorised as: (1) ground-glass opacities; (2) patchy shadows; (3) lobar or segmental consolidation; (4) calcification and/or fibrous streaks; (5) pleural effusion; (6) reticular or honeycombing patterns. Discrepancies were adjudicated by a third senior consultant radiologist.

Treatment information and response assessment

Initial induction remission regimens were recorded, including glucocorticoid type, starting daily dose (converted to prednisone equivalent dose, mg/day), and use of biologics (e.g., mepolizumab, omalizumab) with timing. In our respiratory department, the mainstream immunomodulatory strategy for EGPA during the study period consisted of systemic glucocorticoids with or without anti-IL-5/IL-5R biologics (mepolizumab). Conventional immunosuppressants-including cyclophosphamide, azathioprine, methotrexate, mycophenolate mofetil, and rituximab-were used in only a very small number of patients (n = 7, 3.4%), typically for ANCA-positive, organ-threatening vasculitis following multidisciplinary consultation with rheumatology. Because many patients in this referral centre cohort had received glucocorticoids at varying doses and durations at other hospitals before enrolment, the recorded baseline dose reflects their regimen at the time of first assessment in our department. This may differ from the dose at true disease onset and may partially explain the modest median baseline dose relative to disease activity scores. Pre-referral treatment details were often incompletely documented and could not be systematically analysed.

Treatment response was assessed at 6 months after treatment initiation, following EULAR-defined response criteria for vasculitis [25]: (1) Complete remission: BVAS score of 0 and glucocorticoid dose ≤ 7.5 mg/day prednisone equivalent. (2) Partial remission: ≥50% reduction in BVAS score from baseline but still > 0, or requiring > 7.5 mg/day glucocorticoids to control symptoms. (3) No remission or relapse: failure to achieve partial remission, or recurrence of disease activity after remission.

Statistical analysis and machine learning modelling

All analyses were performed using R language (version 4.2.1) and relevant packages. Continuous variables with normal distribution are presented as mean±standard deviation (SD), compared using one-way ANOVA with Tukey’s HSD post-hoc test; non-normally distributed variables are presented as median (interquartile range, IQR), compared using Kruskal-Wallis test with Dunn’s post-hoc test. Categorical variables are presented as frequency (percentage), compared using Chi-square or Fisher’s exact test (when expected frequency < 5). All tests were two-sided, with p < 0.05 considered statistically significant.

Endophenotype Identification (unsupervised learning)

A strategy combining consensus clustering with the K-means algorithm was employed.

Feature selection and preprocessing

Based on clinical relevance and data completeness, 14 core variables were selected for clustering analysis. All continuous variables were Z-score standardised prior to clustering; missing values were imputed using the median.

Consensus clustering

Performed using the R package “ConsensusClusterPlus” [26]. Parameters were set: 10,000 resampling iterations with replacement, sampling 80% of patients and 80% of features each time. Consensus matrices and cumulative distribution function (CDF) curves for predefined cluster numbers (k) ranging from 2 to 6 were calculated. The optimal k was determined by comprehensive evaluation of: (i) clarity and separation of diagonal blocks in consensus matrix heatmaps; (ii) shape of CDF curves and the incremental change in area under the curve; (iii) consensus scores for different k values.

K-means clustering validation

Based on the optimal k, K-means clustering was performed on all samples (nstart = 25). The silhouette coefficient was used to assess clustering quality, measuring both within-cluster cohesion and between-cluster separation [27].

Dimensionality reduction and visualisation

Principal component analysis was applied to the 14-dimensional feature set for dimensionality reduction. Samples were visualised on a 2D scatter plot using the first two principal components, coloured by their cluster assignment.

Subgroup characterisation and treatment response comparison

Patients were divided into subgroups based on final cluster labels. Systematic comparisons were made across all collected variables. Differences in demographics, clinical features, laboratory markers, imaging patterns, initial treatment, and 6-month treatment response were analysed.

Sensitivity analysis

Two sensitivity analyses were conducted to assess robustness: (i) repeating the consensus clustering and K-means procedure after removing the variable (“FeNO”) with the most missing data; (ii) repeating consensus clustering using Manhattan distance instead of Euclidean distance. The adjusted Rand index (ARI) was used to compare consistency between results from different analyses and the primary analysis (ARI > 0.75 indicates excellent agreement) [28].

Online prediction tool development and validation

A web-based online prediction system for EGPA endophenotypes was developed based on the identified cluster centroids, implementing a weighted Euclidean distance K-nearest neighbours algorithm. Performance was validated in an independent cohort.

Model transformation and deployment

The characteristic vectors of the three identified endophenotype cluster centroids were translated into a computationally applicable predictive model. A k-nearest neighbors classification algorithm was implemented using the scikit-learn library (version 1.2.2) in Python 3.9, with the cluster centroids serving as reference points. The model calculates the weighted Euclidean distance between a patient’s data point and each cluster centroid to assign the phenotype.

System architecture

1. Frontend Interface: A responsive user interface was constructed using HTML5, CSS3, and JavaScript (React.js 18.2.0).

2. Backend Service: The Flask 2.3.2 framework was employed to provide a RESTful API.

3. Computational Engine: The trained model was serialized in joblib format for API calls.

4. Deployment Environment: The tool was containerized using Docker and deployed on a hospital intranet server (CentOS 7.9).

Feature standardization and missing data handling

All continuous variables were standardized using Z-score transformation prior to prediction, utilizing the mean and standard deviation from the original study cohort. For handling missing data, the system offers three strategies: (1) Reject prediction and request complete data entry; (2) Impute missing values using the median of the feature; (3) Employ multiple imputation based on the k-nearest neighbors algorithm.

Tool validation

The tool’s performance was prospectively evaluated in a temporally independent internal validation cohort (n = 30, newly diagnosed patients from January to March 2024). It achieved a prediction accuracy of 90.0% (27/30) and demonstrated excellent agreement with expert clinical phenotyping, as indicated by a Kappa value of 0.850 (95% CI: 0.765–0.935). Detailed performance metrics are presented in Table 4.

Table 4.

Performance validation results of the EGPA endophenotype online prediction tool

Performance Metric	Leave-One-Out Cross-Validation (n = 205)	Prospective Validation (n = 30)	Description
Overall Accuracy	93.7% (192/205)	90.0% (27/30)	Proportion of correctly predicted phenotypes matching original cluster labels.
Sensitivity (by phenotype)
▪ Eosinophilic-Airway Inflammation	96.2% (75/78)	91.7% (11/12)	Ability to correctly identify this phenotype.
▪ ANCA-Systemic Vasculitis	92.3% (60/65)	88.9% (8/9)	Ability to correctly identify this phenotype.
▪ Paucieosinophilic-Multi-Organ Damage	91.9% (57/62)	88.9% (8/9)	Ability to correctly identify this phenotype.
Specificity (by phenotype)
▪ Eosinophilic-Airway Inflammation	95.3% (121/127)	94.4% (17/18)	Ability to correctly exclude this phenotype.
▪ ANCA-Systemic Vasculitis	97.1% (136/140)	95.2% (20/21)	Ability to correctly exclude this phenotype.
▪ Paucieosinophilic-Multi-Organ Damage	96.5% (138/143)	95.2% (20/21)	Ability to correctly exclude this phenotype.
Weighted-average F1 Score	0.937	0.898	Multi-class extension of F1 score, weighted by class support.
Matthews Correlation Coefficient (macro)	0.905	0.850	Multi-class extension of MCC.
Agreement with Expert Judgement
▪ Kappa Value	0.904 (95%CI: 0.867–0.941)	0.850 (95%CI: 0.765–0.935)	Landis-Koch criteria: almost perfect agreement.
Prediction Confidence Distribution
▪ High Confidence (> 80%)	88.3% (181/205)	83.3% (25/30)	Tool’s self-assessed prediction certainty.
▪ Moderate confidence (60–80%)	10.2% (21/205)	13.3% (4/30)
▪ Low confidence (< 60%)	1.5% (3/205)	3.3% (1/30)	Usually prompts re‑checking of input data.
Computational Performance
▪ Average Response Time	< 500 ms	< 500 ms	Time from submission to result return.
▪ Feature‑Missing Tolerance	Supports ≤ 3 missing items	Supports ≤ 3 missing items	Uses k‑nearest‑neighbors imputation for missing data.

In the prospective validation cohort, expert clinical phenotyping was performed independently by two senior rheumatologists, with a third expert arbitrating in cases of disagreement

Results

Patient baseline characteristics

A total of 205 EGPA patients meeting the criteria were included. The mean age was 52.8 ± 13.5 years, with 121 (59.0%) females. All patients had respiratory system involvement, with asthma present in 185 (90.2%) and allergic rhinitis in 147 (71.7%). Other common organ involvements included peripheral neuropathy (67, 32.7%), skin/mucosal lesions (45, 22.0%), gastrointestinal involvement (29, 14.1%), and cardiac involvement (16, 7.8%). Laboratory tests showed a median baseline peripheral blood absolute eosinophil count of 2.34 × 10⁹/L (IQR: 1.03–4.27), median FeNO of 42 ppb (IQR: 17–85), and median serum total IgE of 356 IU/mL (IQR: 158–612). CRP levels were elevated with a median of 22.5 mg/L (IQR: 12.8–35.0). ANCA positivity was found in 23 patients (11.2%), all being MPO-ANCA positive. The median baseline BVAS-3 score was 12 (IQR: 9–15). Detailed baseline characteristics are presented in Table 1. Regarding treatment patterns, the immunomodulatory strategy in this respiratory-based cohort predominantly consisted of systemic glucocorticoids with or without anti-IL-5 biologics (mepolizumab). Conventional immunosuppressants (cyclophosphamide, azathioprine, methotrexate, mycophenolate mofetil, or rituximab) were used in only 7 patients (3.4%).

Table 1.

Baseline characteristics of 205 EGPA patients

Characteristic	Overall (n=205)
Demographics
Age (years, mean±SD)	52.8 ± 13.5
Female (n, %)	121 (59.0)
Clinical Manifestations [n (%)]
Asthma	185 (90.2)
Rhinitis	147 (71.7)
Peripheral Neuropathy	67 (32.7)
Skin/Mucosal Lesions	45 (22.0)
Gastrointestinal Involvement	29 (14.1)
Cardiac Involvement	16 (7.8)
BVAS (median, IQR)	12 (9-15)
Laboratory Investigations
EOS (x10⁹/L)	2.34 (1.03-4.27)
EOS (%)	18.5 (9.2-30.8)
FeNO (ppb)	42 (17-85)
Serum IgE (IU/mL)	356 (158-612)
CRP (mg/L)	22.5 (12.8-35.0)
ANCA Positive (n, %)	23 (11.2)
Imaging Features [n (%)]
Ground-Glass Opacities	134 (65.4)
Consolidation	129 (62.9)
Fibrous Streaks/Calcification	98 (47.8)
Pleural Effusion	28 (13.7)

Abbreviations: EOS Eosinophil count, FeNO Fractional exhaled nitric oxide, BVAS Birmingham Vasculitis Activity Score, IQR Interquartile Range (P25 - P75)

Identification of EGPA Endophenotypes by consensus clustering

Consensus clustering analysis on the 14 core features from 205 patients evaluated model stability for k values from 2 to 6. The optimal number of clusters was determined to be 3 (k = 3). At k = 3, the average silhouette coefficient reached its peak (0.959), significantly higher than for other k values (k = 2: 0.959; k = 4: 0.949; k = 5: 0.955; k = 6: 0.942), indicating optimal balance between within-cluster homogeneity and between-cluster separation. The consensus matrix heatmap clearly showed three distinct, internally consistent diagonal blocks, visually confirming the existence of three stable patient subgroups. The cumulative distribution function curves reached a plateau after k = 3, with minimal incremental area change, suggesting no substantial improvement in consensus stability with more clusters (Fig. 1).

Fig. 1.

Determination of the optimal cluster number by consensus clustering analysis (k = 3). A Consensus cumulative distribution function (CDF) curves for k values 2–6. The shaded area indicates the magnitude of CDF change, which plateaus after k = 3, suggesting no significant improvement in model stability with additional clusters. B Average silhouette coefficients for different k values. The highest coefficient (0.959) at k = 3 indicates the most robust cluster structure. C Consensus matrix heatmap for k = 3. Red indicates pairs of samples more frequently assigned to the same cluster across resampling iterations, while white indicates less frequent co-assignment. Three clear, distinct diagonal blocks corresponding to the three endophenotypes are evident

Subgroup visualisation and endophenotype nomenclature

Based on the K-means clustering results, all patients were assigned to three subgroups. Dimensionality reduction visualisation via principal component analysis (Fig. 2) showed clear separation trends among the three subgroups in the two-dimensional space defined by the first two principal components (cumulative variance explained: 22.27%; PC1: 14.17%, PC2: 8.11%), despite some overlap reflecting the continuous nature of EGPA manifestations. Based on their most prominent clinical and biological features, the subgroups were named:

1. Subgroup 1 (n = 78, 38.0%): Eosinophilic-Airway Inflammation Type

2. Subgroup 2 (n = 65, 31.7%): ANCA-Systemic Vasculitis Type

3. Subgroup 3 (n = 62, 30.2%): Paucieosinophilic-Multi-Organ Damage Type

Fig. 2.

Distribution of EGPA patients in the space of the first two principal components, coloured by endophenotype. Each point represents one patient, coloured by assigned subgroup (blue: Subgroup 1 - Eosinophilic-Airway Inflammation Type; red: Subgroup 2 - ANCA-Systemic Vasculitis Type; green: Subgroup 3 - Paucieosinophilic-Multi-Organ Damage Type). The first two principal components cumulatively explain 22.27% of the total variance (PC1: 14.17%, PC2: 8.11%). The three subgroups show distinct separation in this 2D space, supporting the validity of the clustering results

Clinical-biological characteristics of each endophenotype

The three subgroups showed no significant differences in age or sex distribution (p > 0.05) but exhibited high heterogeneity in nearly all other clinical, laboratory, and imaging characteristics (Table 2; Fig. 3).

Table 2.

Comparison of baseline characteristics among the three EGPA endophenotypes

Characteristic	Overall (n = 205)	Subgroup 1 (n = 78)	Subgroup 2 (n = 65)	Subgroup 3 (n = 62)	p-value
Demographics
Age (years, mean ± SD)	52.8 ± 13.5	50.2 ± 12.8	55.1 ± 14.0	53.3 ± 13.5	0.089
Female (n, %)	121 (59.0)	48 (61.5)	37 (56.9)	36 (58.1)	0.843
Clinical Manifestations (n, %)
Asthma	185 (90.2)	78 (100.0)	60 (92.3)	48 (77.4)	< 0.001
Rhinitis	147 (71.7)	70 (89.7)	45 (69.2)	32 (51.6)	< 0.001
Peripheral Neuropathy	67 (32.7)	10 (12.8)	50 (76.9)	7 (11.3)	< 0.001
Skin/Mucosal Lesions	45 (22.0)	8 (10.3)	24 (36.9)	13 (21.0)	0.001
Gastrointestinal Involvement	29 (14.1)	5 (6.4)	8 (12.3)	18 (29.0)	0.001
Cardiac Involvement	16 (7.8)	3 (3.8)	5 (7.7)	9 (14.5)	0.068
BVAS (median, IQR)	12 (9–15)	11 (8–14)	15 (12–18)	11 (8–14)	< 0.001
Laboratory Investigations
EOS (x10⁹/L)	2.34 (1.03–4.27)	6.10 (5.10−8.00)	1.90 (1.00−2.80)	1.00 (0.60–1.50)	< 0.001
EOS (%)	18.5 (9.2–30.8)	32.0 (28.8–41.0)	15.0 (9.0–23.0)	7.5 (4.5–11.0)	< 0.001
FeNO (ppb)	42 (17–85)	79 (55–112)	38 (12–62)	16 (10–24)	< 0.001
Serum IgE (IU/mL)	356 (158–612)	798 (602–980)	155 (85–248)	128 (60–198)	< 0.001
CRP (mg/L)	22.5 (12.8–35.0)	16.0 (9.0–25.0)	33.5 (22.5–48.5)	18.5 (10.0–29.0)	< 0.001
ANCA Positive (n, %)	23 (11.2)	3 (3.8)	17 (26.2)	3 (4.8)	< 0.001
Imaging Features [n (%)]
Ground-Glass Opacities	134 (65.4)	68 (87.2)	35 (53.8)	31 (50.0)	< 0.001
Consolidation	129 (62.9)	62 (79.5)	39 (60.0)	28 (45.2)	< 0.001
Fibrous Streaks/Calcification	98 (47.8)	25 (32.1)	28 (43.1)	41 (66.1)	< 0.001
Pleural Effusion	28 (13.7)	6 (7.7)	16 (24.6)	6 (9.7)	0.008

Abbreviations: EOS Eosinophil count, FeNO fractional exhaled nitric oxide, BVAS Birmingham Vasculitis Activity Score, IQR Interquartile Range (P25 - P75)

Fig. 3.

Radar chart comparing the three EGPA endophenotypes across five key dimensions. This radar chart visually compares the relative profiles of the three endophenotypes across five key clinical dimensions: BVAS score (reflecting systemic disease activity), peripheral blood absolute eosinophil count (EOS), serum total IgE level, ANCA positivity rate, and asthma history. Subgroup 1 shows prominent outward expansion in EOS, IgE, and asthma dimensions. Subgroup 2 excels in BVAS score and ANCA positivity. Subgroup 3 scores relatively lower across all five indicators, particularly eosinophil-related ones, clearly reflecting its “paucieosinophilic” core feature

1. Subgroup 1: Eosinophilic-Airway Inflammation Type.

This subgroup was characterized by an intense type 2 immuno-inflammatory response. Patients had the highest median absolute eosinophil count [6.10 (IQR: 5.10-8.00) ×10⁹/L], eosinophil percentage [32.0% (IQR: 28.8%-41.0%)], FeNO level [79 (IQR: 55–112) ppb], and serum total IgE concentration [798 (IQR: 602–980) IU/mL], all significantly higher than the other two subgroups (all p < 0.001). Respiratory symptoms were extremely prominent, with asthma prevalence at 100% and rhinitis at 89.7%. Chest HRCT predominantly showed active inflammatory changes, with the highest incidence of ground-glass opacities (87.2%) and consolidation (79.5%). This group had the lowest ANCA positivity rate (3.8%).

• (2)Subgroup 2: ANCA-Systemic Vasculitis Type.

The hallmark of this subgroup was high ANCA positivity and high systemic disease activity. The ANCA positivity rate was 26.2% (17 of 65 patients), significantly higher than other groups (p < 0.001). BVAS score [15 (IQR: 12–18)] was the highest among the three groups, indicating severe systemic vasculitis activity. Clinical manifestations were dominated by multi-system vasculitic involvement, with particularly high rates of peripheral neuropathy (76.9%) and skin/mucosal lesions (36.9%). Laboratory tests showed significantly elevated CRP levels [33.5 (IQR: 22.5–48.5) mg/L, p < 0.001], while eosinophil-related indices were only mildly elevated. Chest HRCT findings were the most diverse.

• (3) Subgroup 3: Paucieosinophilic-Multi-Organ Damage Type.

This subgroup was characterized by relatively low eosinophilic inflammatory markers but more extensive non-respiratory organ damage. EOS [1.00 (IQR: 0.60–1.50) ×10⁹/L], FeNO [16 (IQR: 10–24) ppb], and IgE [128 (IQR: 60–198) IU/mL] levels were the lowest among the three groups (all p < 0.001). Although asthma prevalence (77.4%) remained high, it was significantly lower than in Subgroup 1. This group had the highest proportions of gastrointestinal (29.0%) and cardiac (14.5%) involvement. Imaging revealed a predominance of findings suggestive of chronic or structural change, with calcification and/or fibrous streaks being most common (66.1%).

Differences in treatment strategies and short-term treatment response

Initial treatment strategies differed significantly among the subgroups (Table 3). The proportion of patients receiving mepolizumab was numerically highest in Subgroup 1 (70.5%). Subgroup 2 received higher initial glucocorticoid doses. Conventional immunosuppressants were used in only 7 patients (3.4% overall), predominantly in Subgroup 2 (n = 5, 7.7%), reflecting the more vasculitic phenotype of this subgroup and the involvement of rheumatology consultation. The very low overall usage precluded meaningful statistical comparison of immunosuppressant-based treatment strategies among subgroups. At the 6-month assessment, the observed overall remission rates (complete + partial) were 94.9% (Subgroup 1), 86.2% (Subgroup 2), and 80.7% (Subgroup 3). The rate of no remission/relapse differed significantly among the three groups (5.1% vs. 13.8% vs. 19.4% for Subgroups 1, 2, and 3, respectively; p = 0.012 by Chi-square test), with Subgroup 3 having the numerically highest rate. However, post-hoc pairwise comparisons were not performed due to limited statistical power in subgroup analyses.

Table 3.

Comparison of treatment and short-term response among the three EGPA endophenotypes

Item	Overall (n = 205)	Subgroup 1 (n = 78)	Subgroup 2 (n = 65)	Subgroup 3 (n = 62)	p-value
Baseline Treatment
Initial Glucocorticoid Dose (mg/day), median (IQR)	30 (20–50)	30 (25–50)	40 (30–60)	25 (15–40)	0.003
Use of Mepolizumab (n, %)	85 (41.5)	55 (70.5)	18 (27.7)	12 (19.4)	< 0.001
Use of Omalizumab (n, %)	32 (15.6)	18 (23.1)	5 (7.7)	9 (14.5)	0.056
Use of conventional immunosuppressants (any), n (%)	7 (3.4)	1 (1.3)	5 (7.7)	1 (1.6)	0.072
Treatment Response at 6 Months (n, %)
Complete Remission	68 (33.2)	30 (38.5)	18 (27.7)	20 (32.3)	0.412
Partial Remission	118 (57.6)	44 (56.4)	38 (58.5)	30 (48.4)	0.471
No Remission/Relapse	19 (9.3)	4 (5.1)	9 (13.8)	12 (19.4)	0.012
Overall Remission (CR + PR)	186 (90.7)	74 (94.9)	56 (86.2)	50 (80.7)	0.035

Continuous variable p-value from Kruskal-Wallis test; categorical variables from Chi-square or Fisher’s exact test

Abbreviations: IQR Interquartile Range (P25 - P75)

Sensitivity analysis

Two sensitivity analyses were performed to validate the robustness of the clustering results. First, repeating the entire clustering process after removing the “FeNO” feature (which had relatively more missing data) yielded results highly consistent with the primary analysis (adjusted Rand index, ARI = 0.91). Second, changing the distance metric in consensus clustering from Euclidean to Manhattan distance still identified the optimal cluster number as 3, with excellent agreement with the original typing (ARI = 0.88). These results indicate that the three identified EGPA endophenotypes are robust and not driven by specific analytical methods or the missing data pattern of individual variables.

Development and performance validation of the online prediction tool

To facilitate immediate clinical translation of our findings, we developed a web-based online prediction tool for EGPA endophenotypes, encapsulating the complex machine learning model into an intuitive clinical decision support interface.

Tool development methodology

The tool employs a weighted Euclidean distance algorithm to compute the distance between input patient data and the three identified endophenotype cluster centroids from this study. Prediction probabilities are calculated using the softmax function.

Performance validation results

Tool performance was evaluated using leave-one-out cross-validation on the original cohort and prospective testing on an independent validation set (n = 30, newly diagnosed patients from January-March 2024). Performance metrics are summarised in Table 4.

User interface and functionality

The tool provides the following core functionalities (Fig. 4): (1) Interactive input interface for the 14 core features; (2) Real-time prediction calculation displaying results and confidence levels upon submission; (3) Visualisation outputs including probability bar charts for phenotypes and patient feature radar charts; (4) Clinical decision support offering individualised treatment suggestions based on the predicted phenotype; (5) Report generation exporting a PDF assessment report containing input data, prediction results, and therapeutic recommendations.

Fig. 4.

Schematic of the system architecture and workflow of the EGPA endophenotype online prediction tool. The figure illustrates the overall architecture of the EGPA endophenotype prediction tool, encompassing the user workflow layer and the interface layout layer. The User Workflow Layer describes the complete process from data input (core clinical features), real-time processing, to result output (predicted endophenotype and confidence level) and clinical support (individualized treatment recommendations). The Interface Layout Layer presents the functional modules of the user interface, including the input/output panel, visualization charts, and report generation

Discussion

This study is the first to utilise unsupervised machine learning on multidimensional real-world clinical data to identify and validate three stable and distinct clinical-biological endophenotypes in EGPA: Eosinophilic-Airway Inflammation Type, ANCA-Systemic Vasculitis Type, and Paucieosinophilic-Multi-Organ Damage Type. This framework not only objectively confirms the extreme heterogeneity of EGPA presentations but, more importantly, reveals likely stable patient subgroups driven by different immuno-inflammatory pathways, providing strong data-driven support for moving beyond the traditional ANCA binary classification towards pathophysiology-based precision diagnosis and therapy.

Eosinophilic-airway inflammation type: a type 2 immunity-dominated subgroup and implications for precision targeting

The identified Eosinophilic-Airway Inflammation Type (38.0%) aligns closely with the widely described “eosinophil-driven” EGPA phenotype in the literature [29, 30]. This subgroup exhibits hallmark features of a robust type 2 immune response, extremely high peripheral blood eosinophils, FeNO, and IgE, strongly associated with nearly ubiquitous airway hyperresponsiveness (asthma, rhinitis) and typical pulmonary exudative lesions (ground-glass opacities, consolidation). This strongly implicates IL-5, IL-4/IL-13, and related type 2 cytokine pathways as central to its pathogenesis [15]. In this exploratory analysis, our finding of an observed excellent treatment response associated with the anti-IL-5 monoclonal antibody mepolizumab in this subgroup is consistent with subgroup analyses from pivotal trials like MIRRA and MUSE, which showed greater benefit from mepolizumab in EGPA patients with high baseline eosinophil counts [31, 32]. The 2025 Multidisciplinary Expert Consensus on the Diagnosis and Treatment of EGPA also explicitly states that for EGPA patients primarily characterized by “eosinophilic inflammation,” biologics targeting the IL-5/IL-5R pathway should be considered important therapeutic options, especially for steroid-dependent patients or those with glucocorticoid contraindications [33]. This study provides exploratory real-world evidence that is consistent with the strategy of early identification and targeted treatment of this subgroup, although prospective endophenotype-stratified trials are needed to confirm these observations.

ANCA-systemic vasculitis type: the core of anca-mediated vasculitic pathology and the need for intensive immunosuppression

The discovery of the ANCA-Systemic Vasculitis Type (31.7%) successfully delineates the classic ANCA-mediated necrotising small-vessel vasculitis pathology within the heterogeneous EGPA entity. Characterized by the highest ANCA positivity rate among the three subgroups (26.2%, 17 of 65 patients), marked systemic inflammation (elevated CRP), prominent multi-system vasculitic manifestations (especially peripheral neuropathy, 76.9%), and high BVAS scores, this subgroup’s clinical spectrum more closely resembles microscopic polyangiitis (MPA) [34, 35]. Although the ANCA positivity rate in this subgroup was not as high as that typically reported in microscopic polyangiitis, it was substantially and significantly greater than that in the other two subgroups (3.8% and 4.8%, respectively; p < 0.001), suggesting that ANCA-mediated mechanisms play a predominant role in driving the vasculitic manifestations observed in these patients. This finding supports the existence of an independent subgroup within EGPA driven by ANCA, relying more on immune pathways involving B-cell activation and neutrophil extracellular trap (NET) formation [36, 37]. Here, ANCA likely directly contributes to vascular endothelial damage and necrotising vasculitis via neutrophil activation. Therefore, treatment strategies for this phenotype may need to focus more on controlling the vasculitic process itself. Intensive immunosuppressive regimens based on glucocorticoids combined with cyclophosphamide or rituximab have been recommended for severe ANCA-associated vasculitis by current international vasculitis management guidelines [23, 33]. In our study, the ANCA-Systemic Vasculitis Type received the highest initial glucocorticoid doses and had numerically (though not statistically significantly) greater use of conventional immunosuppressants (7.7% vs. 1.3% and 1.6% in the other subgroups). However, it is important to note that, in this respiratory-based cohort, conventional immunosuppressants were used in only 3.4% of patients overall, reflecting real-world practice patterns in a pulmonary setting where aggressive vasculitis-directed immunosuppression is less commonly employed. This limits our ability to draw conclusions regarding the optimal immunosuppressive strategy for this subgroup, and future studies in vasculitis-centre populations are needed to address this question.

Paucieosinophilic-multi-organ damage type: unmet clinical need, potential novel mechanisms, and therapeutic challenges

The Paucieosinophilic-Multi-Organ Damage Type (30.2%) is the most clinically revealing. It challenges the traditional notion that EGPA is invariably associated with significant peripheral blood eosinophilia. Characterized by relatively low eosinophil-related indices but higher frequencies of severe involvement in critical organs like the heart and gastrointestinal tract, this phenotype may represent a distinct pathological subtype of EGPA. Notably, the identification of this paucieosinophilic subgroup within a cohort that was partly recruited based on unexplained peripheral blood eosinophilia (typically > 1.5 × 10⁹/L) further underscores the robustness of this finding. This recruitment pattern would be expected to enrich the cohort for patients with elevated eosinophil counts; the fact that a substantial subgroup with low eosinophil-related indices was nonetheless identified suggests that this phenotype may be even more prevalent in unselected EGPA populations where eosinophilia is not a prerequisite for referral. Its drivers may not rely entirely on classic type 2 pathways but could involve non-type 2 cytokines (e.g., IL-17, IFN-γ mediated Th1/Th17 responses), activation of tissue-resident eosinophils (whose state may not be fully reflected by peripheral counts), or dysfunction of other myeloid cells (e.g., macrophages) and pro-fibrotic mechanisms [38, 39]. Recent studies suggest potential roles for the Th17/IL-23 pathway, tissue-resident memory T cells, or macrophage-mediated fibrosis in this phenotype [40, 41]. In our study, this subgroup showed the numerically lowest overall observed response to current therapies (both conventional immunosuppressants and anti-IL-5 biologics) and the highest rate of non-response/relapse, highlighting the therapeutic challenges in this subgroup and the limitations of existing strategies. The 2025 expert consensus also emphasises the need for multidisciplinary consultation and exploration of novel mechanism-based therapeutic targets [33]. However, it is important to note that, in the absence of systematic disease duration and VDI data, we cannot fully exclude the possibility that some of the multi-organ manifestations observed in this subgroup-particularly cardiac and gastrointestinal involvement with fibrotic changes on imaging-represent, at least in part, chronic cumulative damage from longstanding disease rather than a distinct active immunopathological endophenotype. This distinction has important therapeutic implications, as interventions targeting active inflammation may be less effective in patients with predominantly irreversible structural organ damage. Furthermore, enhanced monitoring and early intervention for cardiac, gastrointestinal, and other critical organ involvement in this subgroup are paramount.

Methodological strengths, clinical translation, and limitations

Methodologically, the use of consensus clustering, a robust unsupervised learning approach, effectively minimised bias from researcher-predefined classification criteria. The reliability and stability of the typing were validated from multiple angles through PCA visualisation and rigorous sensitivity analysis (ARI > 0.88). Our findings align with emerging trends in EGPA endophenotype research internationally. For instance, a European multicentre study also identified “vasculitis-dominant” and “eosinophil-dominant” types [40]; however, our study further distinguished the Paucieosinophilic-Multi-Organ Damage Type, likely due to the inclusion of more comprehensive clinical features (especially detailed organ involvement patterns), imaging details, and treatment response data, revealing richer disease heterogeneity.

Importantly, because treatment allocation in this real-world study was at the discretion of the treating physician and not randomized, the observed associations between endophenotype and treatment response should be considered exploratory and hypothesis-generating. Confounding by indication-whereby patients with more severe or distinct disease manifestations were more likely to receive certain biologics or higher steroid doses-cannot be excluded. Furthermore, incomplete data on pre-referral treatments, concomitant medication changes, and relapse patterns limit our ability to draw causal inferences. Dedicated prospective, endophenotype-stratified interventional studies are required before any treatment recommendation based on endophenotype can be made.

The endophenotype framework has direct clinical translation potential. Using readily available routine indicators, EOS, FeNO, ANCA, CRP, and key organ (nerve, skin, gastrointestinal, cardiac) involvement, clinicians can perform rapid, low-cost bedside screening for endophenotypes. This facilitates risk stratification at diagnosis and may guide initial treatment decisions in an exploratory manner: considering early biologic introduction for the Eosinophilic-Airway Inflammation Type; considering more intensive immunosuppression or early rheumatology co-management for the ANCA-Systemic Vasculitis Type; and strengthening organ function monitoring while actively seeking multidisciplinary collaboration and novel therapeutic opportunities for the Paucieosinophilic-Multi-Organ Damage Type. The developed online prediction tool (accuracy 90.0%, Kappa 0.850) provides convenient technical support for this translation.

This study has several limitations. First, as a single-centre retrospective study, potential selection bias (e.g., possibly higher proportions of severe or referred cases) exists, and the external validity of the conclusions requires further validation in prospective, multicentre, large cohorts across different regions and populations. Second, due to the real-world nature of data collection in a respiratory department, standardized vasculitis-specific instruments-including the Vasculitis Damage Index (VDI) and the Five-Factor Score (FFS)-were not systematically documented in our electronic medical records. The absence of disease duration and VDI data is particularly relevant for the interpretation of the Paucieosinophilic-Multi-Organ Damage Type, as we cannot fully differentiate active disease from chronic cumulative damage. Some of the multi-organ manifestations observed in this subgroup-particularly cardiac and gastrointestinal involvement with fibrotic changes on imaging-may represent, at least in part, irreversible structural organ damage from longstanding disease rather than a distinct active immunopathological endophenotype. This distinction has important therapeutic implications, as interventions targeting active inflammation may be less effective in patients with predominantly irreversible organ damage. Third, the very limited use of conventional immunosuppressants (3.4% of patients) in this respiratory-based cohort may not reflect practice patterns in dedicated vasculitis centres. This treatment pattern-predominantly glucocorticoids with or without anti-IL-5 biologics-represents both a potential selection bias (as the cohort was enriched for respiratory-predominant and eosinophilic phenotypes) and a limitation in our ability to compare immunosuppressive strategies among endophenotypes. Future studies should compare treatment patterns and outcomes between pulmonary-centred and vasculitis-centred EGPA cohorts. Fourth, because many patients had received glucocorticoids prior to referral, the recorded baseline dose may underestimate true disease severity at onset and pre-referral treatment details were often incompletely documented. Fifth, treatment decisions were not randomized; observed differences in treatment response among subgroups may be partially confounded by differences in initial treatment strategies (confounding by indication). Sixth, the inclusion of some potentially important biomarkers (e.g., serum IL-5, eosinophil cationic protein, eosinophil-derived neurotoxin, specific IgE, histopathological evidence) was limited by the retrospective design, restricting deeper molecular-level mechanistic interpretation of the endophenotypes. Seventh, clustering was based primarily on baseline cross-sectional data; the stability of endophenotypes over the disease course or their potential for transition during treatment interventions was not dynamically assessed. Eighth, although the sample size is relatively large for an EGPA study, it remains moderate, which may affect the statistical power for some subgroup analyses (e.g., cardiac involvement). Ninth, all data originated from a single referral centre, potentially limiting the generalisability of the framework to other healthcare settings or populations.

Future research directions

Based on our findings, future research should include: (1) Validating the generalisability, reproducibility, and prognostic value (e.g., for organ damage accrual, relapse risk, survival) of this tripartite classification in prospective, multicentre, international EGPA registry cohorts, incorporating systematic VDI and FFS assessments; (2) Integrating multi-omics data (genomics, transcriptomics, proteomics, metabolomics) to elucidate the molecular basis and driving mechanisms of each endophenotype, particularly the Paucieosinophilic-Multi-Organ Damage Type; (3) Conducting endophenotype-based prospective intervention studies or randomised controlled trials, e.g., comparing early mepolizumab versus conventional step-up therapy in the Eosinophilic-Airway Inflammation Type, or exploring efficacy of novel target drugs for the Paucieosinophilic type; (4) Utilising longitudinal data to study the dynamic evolution of endophenotypes; and (5) Comparing treatment patterns and outcomes between pulmonary-centred and vasculitis-centred EGPA cohorts to understand how clinical setting influences endophenotype-associated management strategies.

Conclusion

Using a data-driven unsupervised machine learning approach on a real-world EGPA patient cohort, this study successfully identified three stable and clinically distinct clinical-biological endophenotypes. This framework profoundly elucidates the heterogeneous nature of EGPA, advancing the traditional ANCA binary classification to a multifaceted classification based on different pathophysiological mechanisms. The distinct characteristics of each endophenotype in clinical presentation, laboratory profile, imaging patterns, and observed treatment response patterns provide an exploratory real-world foundation for implementing individualised, precision treatment strategies. Furthermore, the developed online prediction tool facilitates rapid translation of research findings into clinical practice.

Based on these results, we recommend integrating the concept of “endophenotype identification” into future EGPA clinical assessment and management pathways. Preliminary classification using routine indicators can help inform early differentiation of treatment strategies, optimise patient outcomes, improve treatment efficiency, and provide more precise direction for drug development and clinical trial design in this field. We advocate for the prospective validation and refinement of this classification in multicentre, multidisciplinary cohorts before its incorporation into management guidelines.

Authors’ contributions

Jiangtao Lin and Xuefeng Ni supervised the study. Bingqing Sun, Zizhong Wang, Yuan Li, Ruiheng Zhao and Kunlu Shen contributed to study design, data analysis, and manuscript drafting. Hui Lian and Yecheng Liu contributed to data analysis and visualization. Jiangtao Lin made critical revisions to the manuscript. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (8217010602).

Data availability

The datasets presented in this study can be found in online repositories.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.