The Florida Scoring System for stratifying children with suspected Sjögren’s disease: a cross-sectional machine learning study

Summary

Background:

Childhood Sjögren’s disease is an underdiagnosed, poorly-understood, and rare condition. This first data-driven study, by integrating machine learning models on a rare pediatric cohort in the U.S., aims to develop a novel system, namely the Florida Scoring System (FSS), for stratifying symptomatic pediatric patients suspected of Sjögren’s disease.

Methods:

A rare cohort of 217 symptomatic patients who visited Pediatric Rheumatology at the University of Florida between January,16, 2018 and April, 28, 2022 underwent comprehensive examinations to rule out/in childhood Sjögren’s disease. Latent class analysis (LCA) utilized clinical and laboratory variables to detect heterogeneous patient classes. Machine learning models, including random forest, gradient-boosted decision tree, partial least square discriminatory analysis, LASSO-penalized ordinal regression, artificial neural network, and Super Learner predicted patient classes and ranked variable importance. Causal graph learning selected key features to build the final FSS.

Findings:

The LCA model identified three distinct patient classes: Class I (Dryness Dominant with Positive Tests; DDPT, n=27), Class II (High Symptoms with Negative Tests; HSNT, n=98), and Class III (Low Symptoms with Negative Tests; LSNT, n=92). Machine learning models accurately predicted patient classes and ranked variable importance consistently. Causal graphical model discovered key features for constructing FSS. We found a highly symptomatic patient group with negative serology and diagnostic profiles, which warrants clinical attention. We further revealed that salivary gland ultrasonography can be a non-invasive alternative to minor salivary gland biopsy in children.

Interpretation:

FSS is pediatrician-friendly tool to assist classification and long-term monitoring of suspected childhood Sjögren’s disease. The resulting stratification have important implications for clinical management, trial design, and pathobiological research. FSS requires validation in large prospective pediatric cohorts.

Introduction

Sjögren’s disease is a chronic, systemic disease caused by autoimmune-induced inflammation of the salivary and lacrimal glands, leading to oral and/or ocular dryness (sicca) and extraglandular symptoms.¹ Currently, children only represent 1% of the Sjögren’s population and are under-represented in Sjögren’s research,^2,3 suggesting poor understanding and possible underdiagnosis of this rare condition. With initial presentations often being recurrent parotitis (RP) and salivary gland swelling,^4,5 childhood Sjögren’s disease is notably different from the adult disease. Extraglandular involvement occurs in approximately 50% of pediatric patients, resulting in neurologic, musculoskeletal, and other systemic manifestations.⁶

Currently, a consensus criterion for diagnosing and classifying childhood Sjögren’s disease is still lacking. The 2002 American-European Consensus Group (AECG) criteria, the 2012 American College of Rheumatology (ACR) classification criteria, and the 2016 ACR/European League Against Rheumatism (EULAR) criteria for primary adult Sjögren’s disease had unacceptably low sensitivity in detecting Sjögren’s in children,^7,8 as they often failed to capture the diverse pediatric manifestations. Other proposed children-specific criteria, relied solely on experts’ opinions,⁹ used under-sampled case report (n=8),¹⁰ or summarized from previous literature findings.⁴

Unprecedentedly in the field, we stratified the University of Florida (UF) childhood Sjögren’s disease cohort, the only prospective cohort available in the U.S. at present by integrating machine- and causal graph learning on 196 clinical, laboratory, and diagnostic variables. The goal of the study was to provide an interpretable and clinically applicable classification system, namely the Florida scoring system (FSS), to improve the detection of childhood Sjögren’s disease, assist clinical decision, facilitate patient triage, and prevent a delay in treatment, especially among patients not fulling the adult criteria due to negative serology, minor salivary gland biopsy (MSGBx), or secretory dysfunction.

Methods

Study design and participants.

In this cross-sectional study, we included 217 symptomatic participants (median age at diagnosis:15 years old, 72% female) from the prospective UF childhood Sjögren’s disease cohort. Between January,16, 2018 and April, 28, 2022, patients were first screened by pediatric rheumatologist at the Pediatric Rheumatology Clinic of the UF Health Shands Hospital for possible childhood Sjögren’s disease, and referred to the Oral Medicine Clinic and the Center for Orphaned Autoimmune Disorders (COAD) for further evaluation. Data used for this study were from first visits. Twenty participants, which were diagnosed later as adults, were included, because their symptom onset was before 18. In the absence of a pediatric criteria, an initial evaluation of childhood Sjögren’s disease was conducted based on the 2016 ACR/EULAR adult Sjögren’s disease criteria, and 50 out of 217 patients fulfilled the criteria.¹¹ Suspected Sjögren’s with other suspected comorbidities were included, as there were no specific guidelines differentiating primary from secondary childhood Sjögren’s disease. A detailed description of the inclusion and exclusion criteria for this study was listed in the Appendix (p2). This study was reviewed and approved by the UF Institutional Review Board (Protocols #201600490 and 201900645) with written informed consents/assents obtained from parents/guardians of patients or patients older than 18 years old.

Collection of demographical, clinical, and laboratory information.

All information and specimens were banked at COAD, where the multi-disciplinary collaboration of providers is available for comprehensive patient care and research. Medical history and clinical symptoms, including the items related to the twelve domains of the physician-reported EULAR- Sjögren’s syndrome (SS)-disease-activity-index (ESSDAI) ¹² and the EULAR-SS-patient-reported-index (ESSPRI) ¹³, were analyzed. Unstimulated salivary flow rate (USFR) ≤ 0.1ml/min was considered positive for hypo-salivation. Tear production ≤ 5 mm by the Schirmer’s test constituted severe dry eyes. Ocular surface staining (OSS) was performed in only a small number of patients and was therefore excluded from the analysis. MSGBx was positive if one or more foci of at least 50 mononuclear immune cells was present per 4 mm² of tissue. Salivary gland ultrasound (SGUS) of the bilateral parotid and submandibular glands was defined positive when the summation score ≥17 and/or the simplified score or the OMERACT score was 2 or 3 ^14–16. Detailed information on procedures for obtaining laboratory and diagnostic results are provided in the Appendix (p2).

Data preprocessing and clinically guided variable selection and reduction.

A total of 196 clinical and laboratory variables were extracted from the COAD bank. A comprehensive description for variables is provided in the Appendix (p2–3). Variables with ≥ 40% missing values were excluded. Correlation analysis was conducted to explore the relationship and detect collinearity between variables (Appendix, p7). For the downstream latent class analysis (LCA), which requires local independence of the included variables within each latent class, variables with Spearman’s rank correlation coefficient absolute values >0.65 were combined (if clinically similar) or excluded from the analysis. A final panel of 33 variables was chosen based on their clinical relevance¹⁷ and importance in evaluating childhood Sjögren’s disease by experts, including hypergammaglobulinemia, cytopenia, autoimmune hemolytic anemia (AIHA), low C3, low C4, anti-SSA, anti-SSB, ANA, RF, anti-dsDNA, anti-RNP, anti-TPO, anti-thyroglobulin, anti-mitochondrial, anti-centromere, eSjA (CA6/SP1/PSP at least one positive), USFR, focus score, Schirmer’s test, SGUS, recurrent parotitis/glandular swelling, dry mouth (subjective), dry eyes (subjective), candidiasis, and ESSPRI- and ESSDAI-related symptoms and signs, entered the final analyses. Depending on the ability of different models to handle missingness, missing values were automatically excluded from the class probability calculation, reconstructed by the generalized low-rank model before k-means clustering, and imputed for other analyses, respectively.

Statistical analyses.

Figure 1 summarized the analytical pipeline of this study. All analyses were performed using R V.4.2.0 (http://www.r-project.org). Latent class analysis (LCA), validations, machine learning models, and FSS are explained herein.

Figure 1. Schematic presentation of the analytical pipeline for the development of FSS.

We first performed latent class analysis (LCA) for the identification of clinically meaningful patient groups, validated the LCA-derived patient classes using 10-fold cross-validation, and compared them with k-means clusters. Supervised machine learning models were applied to predict LCA-derived patient classes and rank variable importance. The causal graph algorithm was applied to discover variables retaining critical information for predicting LCA-derived patient classes, which were used for the construction of the weighted FSS scoring system. RF: random forest, GBDT: gradient-boosted decision tree, PLSDA: partial least square discriminatory analysis. ANN: artificial neural network. LASSO-OR: Least Absolute Shrinkage and Selection Operator penalized ordinal regression.

LCA for identifying heterogeneous patient classes.

To uncover the heterogeneous patient classes, we applied LCA via the poLCA R package. We constructed a series of LCA models using the 33 indicator variables (mentioned above) from all 217 participants, with the number of classes ranging from 2 to 6. The underlying assumption of LCA is that latent classes exist and drive the differences in the distribution of observed variables across the population. The resulting classes can be related to disease status based on class-specific clinical and laboratory patterns, as shown in previous literature including the diagnosing primary Sjögren’s in adults ¹⁸. As poLCA requires polytomous variables, the 11-scale ESSPRI scores (e.g., ESSPRI-Dryness, ESSPRI-Pain, and ESSPRI-Fatigue) were re-categorized into 4 levels: 0 (no), 1–3 (mild), 4–6 (moderate), 7–10 (severe). Model fitness was examined by the Bayesian information criterion (BIC), and the number of classes that minimize the BIC was chosen as optimal. The final clustering results were shared with clinical experts, to further define clinically meaningful classes tailored to the distinct patient profiles. We implemented a 2-stage strategy to internally validate our LCA categorization, which included a 10-fold cross-validation (CV) to assess the stability of class labels and a comparison with another clustering method (k-means). More detailed information on LCA validation is provided in the Appendix (p3). To assess the clinical importance of each indicator variable, we further estimated their associations with the LCA-defined patient classes and the 2016 ACR/EULAR criteria-defined disease groups, using univariable ordinal regression and univariable logistic regressions, respectively (Appendix, p3).

Sensitivity analysis.

To assess the impact of young adult patients who were older than 18 years (n=20) and patients with missingness in critical diagnostic variables (e.g., anti-SSA, USFR, SGUS, and MSGBx; n=107) on the classification results, we re-conducted the LCA, by excluding these patients, respectively.

Supervised machine learning for LCA class prediction.

To investigate the ability of machine learning algorithms for predicting LCA-defined patient classes, we tested the Random Forest, the Gradient Boosted Decision Tree (GBDT), the Least Absolute Shrinkage and Selection Operator-penalized ordinal regression (LASSO-OR), the Partial Least Square Discriminant Analysis (PLS-DA), the Artificial Neural Network (ANN), and the Super Learner models (which stacked most of the above algorithms into a meta-algorithm). The predictors for all models were the 33 clinical and laboratory variables mentioned above, and the outcome was patient classes. The entire sample (n=217) was split into training and testing sets at a 7:3 ratio, wherein the training set (70%) was used for model construction and hyperparameter tuning, and the testing set (30%) was for performance evaluation. Detailed information on hyperparameters tuning, variable importance evaluation, the ANN model, and the Super Learner model is available through the Appendix (p3–4). In brief, Radom Forest, GBDT, LASSO-OR, and PLS-DA were tuned by repeated CV (5-fold, repeated 50 times) on the training dataset. The contribution of each individual algorithm to the meta-algorithm in the Super Learner was assessed using 5-fold CV, while the performance of each individual algorithm and the meta-algorithm on the training set was evaluated using nested CV (inner 5-fold; outer: 5-fold). To further assess the performance of all above models on a separate testing set (30%), we calculated the overall prediction accuracy as well as class-specific sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), balanced accuracy, and area under the receiver operating characteristic curve (AUROC).

Causal graph learning for the construction of FSS.

To discover variables that retained essential information for predicting LCA-defined patient classes, we applied causal graph learning through the R package rCausalMGM.¹⁹ rCausalMGM starts with a mixed graphical model (undirected graph), from which edges between conditionally independent variables were gradually removed. We optimized the hyper-parameter $\text{λ}$ (which controls the sparsity of edges in the undirected graph), using a stability-based method.²⁰ Then we applied fast causal inference (FCI) to estimate conditional independence between variables and learn the final directed graph. The FCI optimal sparsity parameter ($\text{α}$) was selected through 10-fold CV. Further incorporating clinicians’ input, we removed one redundant item and applied ordinal regression on the remaining variables, to derive the weighted scoring system FSS.

Role of the funder.

The funder had no role in data collection, analysis, interpretation, writing the manuscript, or the decision to submit.

Results

Between January 16, 2018, and April 28, 2022, we screened 448 patients. After excluding 205 patients due to their symptom onset later than 18 years of age, we recruited 243 patients into our cohort. Additional 26 patients were excluded because of confirmed diagnosis of systemic lupus erythematosus (SLE), juvenile parotitis, juvenile idiopathic arthritis (JIA), or infection with a clear etiological microorganism, which resulted in a total of 217 patients entering the final analysis. The median patient age at diagnosis were 15 years (interquartile range: 11–17), 155(72%) of patients were female, 167 (79%) were White, and 20 (9%) were Hispanic/Latino/Spanish.

LCA yielded three distinct classes, and across 10-fold CV, this class number was consistently selected, and the mean agreement between original class labels and predicted labels was 0.898 (95% CI: 0.695–0.983), suggesting stable clustering. Based on the class-specific clinical, laboratory, and diagnostic features (Table 1) and to stay consistent with the previous studies^21,22, we defined the patient classes as Class I: Dryness Dominant with Positive Tests (DDPT, n=27); Class II: High Symptoms with Negative Tests (HSNT, n=98); and Class III: Low Symptoms with Negative Tests (LSNT, n=92). Dryness dominant is defined as the class-median ESSPRI-Dryness score ≥ 5 with other ESSPRI class-medians<5; high symptom is defined as class-medians of all three ESSPRI scores ≥5, and low symptom is defined as all three ESSPRI class-medians <5. Most of the Class I patients (70%) fulfilled the 2016 ACR/EULAR criteria, whereas only 18% of Class II and 14% of Class III patients met the criteria. Patient demographic, clinical, and laboratory characteristics by 2016 ACR/EULAR criteria were summarized in Appendix (p13–16), in which we observed no significant differences in subjective symptom measures (including sicca), between patients who met and who did not meet the criteria. By contrast, LCA Class I was more likely to have high erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), complements C3 and C4; positive SGUS; positive autoantibodies; compared to Class II and Class III. The prevalence of hypergammaglobulinemia, cytopenia, autoimmune hemolytic anemia, recurrent parotitis/glandular swelling, SLE-like presentations, immune-related conditions, and renal and cutaneous symptoms was also higher in Class I. Class II had the highest ESSPRI scores, the highest prevalence of sicca symptoms along with articular, neurological, and gastrointestinal involvement. Patients in class III were notably less symptomatic than patients in Class II. However, objective laboratory and diagnostic tests, such as anti-SSA and MSGBx, did not significantly differ between Class II and Class III. Association of each indicator variable (input variables of the LCA model) with the LCA-defined patient classes and the 2016 ACR/EULAR-defined disease groups were illustrated in Appendix (p8), which shows the top three variables with strongest association with LCA-defined patient classes were cytopenia, hypergammaglobulinemia, and anti-SSA.

Table 1.

Demographic, clinical, and laboratory characteristics of the UF pediatric cohort based on the latent class analysis (LCA)^*

	All (n=217)	Class I (DDPT, n=27)	Class II (HSNT, n=98)	Class III (LSNT, n=92)	P †
Patients meeting the 2016 ACR/EULAR criteria (n=50)		19/27 (70%)	18/98 (18%)	13/92 (14%)	<.0001
Age (years), median (IQR)
Age of symptom onset	12 (8, 15)	12 (9, 14)	13 (9, 15)	11 (7, 14)	0.05
Age of diagnosis§	15 (11, 17)	15 (12, 17)	16 (14, 17)	13 (9, 16)	<.0001
Sex, n/N (%)					0.029
Female	155/216 (72%)	19/27 (70%)	78/97 (80%)	58/92 (63%)
Male	61/216 (28%)	8/27 (30%)	19/97 (20%)	34/92 (37%)
Race, n/N (%)					0.0005‡
White	167/212 (79%)	8/27 (30%)	84/95 (88%)	75/90 (83%)
Black/African American	27/212 (13%)	14/27 (52%)	7/95 (8%)	6/90 (7%)
Asian/Pacific Islander	6/212 (3%)	1/27 (4%)	1/95 (1%)	4/90 (4%)
Other	12/212 (6%)	4/27 (15%)	3/95 (3%)	5/90 (6%)
Ethnicity, n/N (%)					0.23‡
Hispanic/Latino/Spanish	20/213 (9%)	5/27 (19%)	8/96 (8%)	7/90 (8%)
Non-Hispanic/Latino/Spanish	193/213 (91%)	22/27 (81%)	88/96 (92%)	83/90 (92%)
Lab parameters, n/N (%)
High ESR	38/213 (18%)	12/26 (46%)	20/97 (21%)	6/90 (7%)	0.0005‡
High CRP high	32/209 (15%)	5/26 (19%)	21/94 (22%)	6/89 (7%)	0.0090‡
Low complement C3	10/194 (5%)	8/26 (31%)	1/92 (1%)	1/76 (1%)	0.0005‡
Low complement C4	24/194 (12%)	10/26 (39%)	8/92 (9%)	6/76 (8%)	0.0010‡
Hypergammaglobulinemia	19/216 (9%)	15/27 (56%)	1/98 (1 %)	3/91 (3%)	0.0005‡
Cytopenia	13/216 (6%)	11/27 (41%)	1/98 (1%)	1/91 (1%)	0.0005‡
Autoimmune hemolytic anemia	7/216 (3%)	5/27 (19%)	1/98 (1%)	1/91 (1%)	0.0010‡
Autoantibodies, n/N (%)
Anti-SSA	36/214 (17%)	22/27 (82%)	8/97 (8%)	6/90 (7%)	0.0005‡
Anti-SSB	23/213 (11%)	11/27 (41%)	5/97 (5%)	7/89 (8%)	0.0005‡
Anti-SSA only	22/212 (10%)	11/27 (41%)	7/96 (7%)	4/89 (4%)	0.0005‡
Anti-SSB only	8/212 (4%)	0/27 (0%)	3/96 (3%)	5/89 (5%)	0.31‡
Anti-SSA and anti-SSB	14/212 (7%)	11/27(41%)	1/96(1%)	2/89 (2%)	0.0005‡
Antinuclear antibodies (ANA)	92/206 (48%)	25/26 (96%)	34/93 (37%)	33/88 (38%)	<.0001
Rheumatic factors (RF)	16/133 (12%)	11/26 (58%)	2/63 (3%)	3/51 (6%)	0.0005‡
Anti-dsDNA	20/162 (12%)	10/26 (39%)	4/72 (6%)	6/64 (9%)	0.0010‡
Anti-ribonucleoprotein (RNP)	14/140 (10%)	10/26 (39%)	2/63 (3%)	2/51 (4%)	0.0005‡
Anti-thyroid peroxidase (TPO)	23/183 (13%)	1/24 (4%)	12/85 (14%)	10/74 (14%)	0.45‡
Anti-thyroglobulin	22/168 (13%)	2/21 (10%)	13/81 (16%)	7/66 (11%)	0.60‡
Anti-mitochondrial	8/140 (6%)	4/19 (21%)	4/61 (7%)	0/60 (0%)	0.0040‡
Anti-centromere	4/147 (3%)	1/19 (5%)	2/67 (3%)	1/61 (2%)	0.78‡
Early Sjögren’s autoantibody (eSjA)**	124/184 (67%)	9/18 (50%)	59/81 (73%)	56/85 (66%)	0.16
Diagnostic tests
Unstimulated salivary flow rate (USFR, ml/min), median (IQR)	0.28 (0.14, 0.47)	0.23 (0.08, 0.47)	0.25 (0.13, 0.41)	0.32 (0.17, 0.54)	0.14
Objective dry Mouth (USFR ≤ 0.1 ml/min), n/N (%)	58/216 (27%)	10/27 (37%)	28/98 (29%)	20/91 (22%)	0.40
Focus score, median (IQR)	2 (1, 4)	4 (0, 8)	0 (0, 1)	1 (0, 2)	0.0006
Minor salivary gland biopsy (MSGBx) positive (Focus score ≥ 1), n/N (%)	81/158 (51%)	15/21 (71%)	35/75 (47%)	31/62 (50%)	0.15
Schirmer’s test positive, n/N (%)	25/143 (19%)	3/14 (21%)	12/66 (18%)	10/54 (19%)	0.96
Salivary gland ultrasound (SGUS) positive, n/N (%)	37/158 (24%)	20/22 (91%)	4/74 (5%)	13/62 (21%)	<.0001
Medical history, n/N (%)
Recurrent parotitis/salivary gland swelling (RP/SG swelling)	71/217 (33%)	15/27 (56%)	27/98 (28%)	29/92 (32%)	0.022
Infection (bacterial, viral, or fungal) ††	59/217 (27%)	10/27 (37%)	29/98 (30%)	20/92 (22%)	0.22
Immune-related conditions††	103/217 (48%)	21/27 (78%)	46/98 (47%)	36/92 (39%)	0.0003
Systemic lupus erythematosus (SLE)-like presentation‡‡	19/216 (9%)	16/27 (59%)	1/98 (1%)	2/90 (2%)	0.0005‡
Chronic arthritis§§	27/216 (13%)	6/27 (22%)	10/98 (11%)	11/91 (11%)	0.24‡
Asthma	71/216 (33%)	4/27 (15%)	39/98(40%)	28/91 (31%)	0.043
Amplified musculoskeletal pain syndrome (AMPS) /fibromyalgia	17/216 (8%)	0/27 (0%)	15/98 (15%)	2/91 (2%)	0.0020‡
Attention deficit hyperactivity disorder (ADHD) /Autism	26/217 (12%)	1/27 (4%)	17/98 (17%)	8/92 (9%)	0.056‡
Clinical features, n/N (%)
Sicca	103/216 (48%)	15/27 (56%)	76/98 (78%)	12/91 (13%)	<.0001
Dysphagia	52/216 (24%)	2/27 (7%)	37/98 (38%)	13/91 (14%)	<.0001
Oral ulcer	60/216 (28%)	9/27 (33%)	34/98 (35%)	17/91 (19%)	0.039
Dental issues	60/216 (28%)	10/27 (37%)	28/98 (29%)	22/91 (24%)	0.41
Dry skin	54/216 (25%)	6/27 (22%)	31/98 (32%)	17/91 (19%)	0.11
Rash	112/216 (52%)	15/27 (56%)	55/98 (56%)	42/91 (46%)	0.36
Migraines/Headache	18/216 (8 %)	18/27 (67%)	86/98 (88%)	56/91 (62%)	0.0020
Dysautonomia	145/216 (67%)	1/27 (4%)	35/98 (36%)	14/91 (15%)	0.0002
Raynaud’s phenomenon	50/216 (23%)	4/27 (15%)	10/98 (10%)	4/91 (4%)	0.15
Temporomandibular joint disorders	41/217 (19%)	3/27 (11%)	28/98 (19%)	10/92 (11%)	0.0042
Hypermobility spectrum disorder/Ehlers-Danlos syndrome (HSD/EDS)	122/216 (57%)	12/27 (44%)	75/98 (77%)	35/91 (39%)	<.0001
Candidiasis	8/216 (4%)	1/27 (4%)	6/98 (6%)	1/91 (1%)	0.19
Parotid gland tenderness (unilateral)	27/217 (12%)	8/27 (30%)	12/98 (12%)	7/92 (8%)	0.010‡
Parotid gland tenderness (bilateral)	39/217 (18%)	4/27 (15%)	23/98 (24%)	12/92 (13%)	0.12‡
Submandibular gland tenderness (unilateral)	6/217 (3%)	0/27 (0%)	5/98 (5%)	1/92 (1%)	0.16‡
Submandibular gland tenderness (bilateral)	14/217 (7%)	0/27 (0%)	9/98 (9%)	5/92 (5%)	0.17‡
Negative parotid gland milking (unilateral)	36/215 (17%)	8/27 (30%)	16/97 (17%)	12/91 (13%)	0.23
Negative parotid gland milking (bilateral)	84/215 (39%)	9/27 (33%)	39/97 (40%)	36/91 (40%)	0.95
Negative submandibular gland milking (bilateral)	59/216 (29%)	6/27 (22%)	30/93 (32%)	23/89 (26%)	0.55
Negative saliva pooling	61/208 (29%)	6/27 (22%)	32/93 (34%)	23/89 (26%)	0.34
ESSDAI-*** and ESSPRI-related signs/symptoms
ESSPRI Dryness, median (IQR)	4 (0, 6)	5 (0, 5)	6 (4, 7)	0 (0, 3)	<.0001
ESSPRI Fatigue, median (IQR)	5 (2, 8)	3 (0, 6)	8 (6, 9)	3 (0, 5)	<.0001
ESSPRI Pain, median (IQR)	3 (0, 6)	0 (0, 6)	5 (3, 7)	0 (0, 3)	<.0001
ESSPRI Mean, median (IQR)	4.0 (2.0, 6.0)	2.7 (1.0, 4.8)	5.7 (4.7, 7.0)	2.0 (0.7, 3.0)	<.0001
Fever, n/N (%)	44/216 (20%)	7/27 (26%)	20/98 (20%)	17/91 (19%)	0.71
Weight loss, n/N (%)	17/216 (8%)	5/27 (19%)	7/98 (7%)	5/91 (6%)	0.077‡
Lymphadenopathy, n/N (%)	30/217 (14%)	5/27 (19%)	8/98 (8%)	17/92 (19%)	0.086‡
Cutaneous, n/N (%)	4/217 (2%)	3/27 (11%)	0/98 (0%)	1/92 (1%)	0.0045‡
Pulmonary, n/N (%)	117/217 (55%)	15/27 (56%)	62/98 (63%)	40/92 (44%)	0.023
Renal, n/N (%)	37/217 (17%)	15/27 (56%)	13/98 (13%)	9/92 (10%)	0.0005‡
Muscular, n/N (%)	96/217 (44%)	10/27 (37%)	62/98 (63%)	24/92 (26%)	<.0001
Articular, n/N (%)	174/217 (80%)	20/27 (74%)	96/98 (98%)	58/92 (63%)	0.0005‡
Neurological, n/N (%)	152/217 (70%)	19/27 (70%)	96/98 (98%)	67/92 (73%)	0.0005‡
Cardiovascular, n/N (%)	41/217 (19%)	5 /27 (19%)	24/98 (25%)	12/92 (13%)	0.13
Gastrointestinal, n/N (%)	129/217 (59%)	10/27 (37%)	81/98 (83%)	38/92 (41%)	<.0001
Other domains†††, n/N (%)	33/217 (15%)	4/27 (14.8%)	24/98 (25%)	5/92 (5%)	0.0030‡

Continuous variables were presented in median (IQR). Categorical variables were presented as n/N (%), where n represented the number of cases, N represented numbers of total patients without missing values, and % represented the percentage of n in N.

^*Classification is based on the 3-class LCA model with 33 variables, including hypergammaglobulinemia, cytopenia, hemolytic anemia, low C3, low C4, Anti-SSA, Anti-SSB, ANA, RF, anti-dsDNA, anti-RNP, anti-TPO, anti-thyroglobulin, anti-mitochondrial, anti-centromere, early SS (SP1/CA6/PSP at least one positive) antibodies, objective dry mouth (unstimulated flow rate, USFR), MSGBx (focus score), Schirmer’s test, salivary gland ultrasound (SGUS), recurrent parotitis (RP)/glandular swelling, dry mouth (subjective), dry eyes (subjective), candidiasis, ESSPRI Dryness, ESSPRI Pain, ESSPRI Fatigue, fever, weight loss, renal, neurological, articular, gastrointestinal. The original 11-scale ESSPRI (Dryness, Pain, and Fatigue) scores were re-categorized into 4 levels: 0 (no), 1–3 (mild), 4–6 (moderate), 7–10 (severe).

^†Calculated by Chi-square’s test or Wilcoxon-Mann-Whitney test, unless the assumption underlying the Chi-square’s test was violated.

^‡When assumption of the Chi-square’s test was violated, i.e., the expected cell count was less than 5 in more than 20% of the cells, the P-value was computed by Monte Carlo simulation with 2000 replications.

^§Age at which childhood Sjögren’s disease were diagnosed or ruled out, following the completion of tests required by the 2016 ACR/EULAR classification criteria for primary Sjögren’s syndrome in adult patients. Of note, a total of 20 participants older than 18 years were included, because they were under the care of pediatric rheumatologists for years since symptom onset before 18.

^**Encompassed nine items, including IgG, IgA, and IgM of anti-SP1, anti-CA6, and anti-PSP, respectively. If at least one of the nine items is positive, eSjA is defined as positive. Otherwise, eSjA is defined as negative.

^††Detailed information of infection and immune-related conditions are available in online supplementary methods.

^‡‡SLE-like presentation: positive symptoms or autoantibodies but not fulfilling the Systemic Lupus International Collaborating Clinics Classification (SLICC) Criteria

^§§Chronic arthritis: symptoms and signs of arthritis lasting over 6 weeks with other symptoms (e.g.,fatigue, dryness) but not fulfilling the International League of Associations for Rheumatology (ILAR) classification criteria for juvenile idiopathic arthritis.

^***The details for positive neurological, articular, and other domains are available in online supplementary methods. As systemic features of Sjögren’s disease in children have not been clearly defined, ESSDAI-related items with additional features, such as cardiovascular or gastrointestinal domains, were all included.

^†††Other domains of ESSDAI include fibromyalgia, amplified pain syndrome, and Raynaud’s phenomenon.

DDPT: Dryness Dominant with Positive Tests

HSNT: High Symptoms with Negative Tests

LSNT: Low Symptoms with Negative Tests

IQR: interquartile range.

ESR: erythrocyte sedimentation rate.

CRP:C-reactive protein.

ESSDAI: physician-reported European League Against Rheumatism (EULAR) - Sjögren’s syndrome (SS)-disease-activity-index.

ESSPRI: European League Against Rheumatism (EULAR) - Sjögren’s syndrome (SS) -patient-reported-index.

Appendix (p9) further compares all symptomatic patient groups defined by different classification criteria and clustering methods. 48% of patients who did not fulfill the 2016 ACR/EULAR criteria were captured by LCA Class II. Patients in k-means clusters 1–2 tend to fall mainly into LCA Class I, clusters 3–4 mainly into LCA Class II, and cluster 6 mainly into LCA Class III, indicating reasonable agreement between two clustering methods. Sensitivity analysis showed that, after removing the young adult patients, the 3-class LCA model had a Cohen’s Kappa coefficient (Κ) of 0.846 (Mcnemar’s P=0.0026) with the original classification, indicating strong agreement.²³ After removing patients with missingness on anti-SSA, USFR, SGUS, or MSGBx, the 3-class LCA model had a Κ of 0.764 (Mcnemar’s P=0.02442) with the original classification, indicating moderate agreement.²³

All computational models had high overall accuracy in predicting LCA-derived patient classes. Super Learner performed the best in differentiating Class I (balanced accuracy=1.000), GBDT best in differentiating Class II (balanced accuracy=0.954, lower limit of 95% CI: 0.903), and PLS-DA best in differentiating Class III (balanced accuracy: 0.974, lower limit of 95% CI: 0.939). Sensitivity, specificity, PPV, NPV, overall accuracy, and AUROC of each model estimated on the testing set were further presented in Appendix (p17 and p10, respectively). The contribution of each individual model to the Super Learner was listed in Appendix (p18). LASSO had the largest contribution for predicting Class I, and Random Forest for predicting Class II and Class III. The performance of each individual model and the meta-model, as evaluated by mean squared error during 5-fold nested CV was listed in Appendix (p19), in which LASSO had the best performance for predicting Class I, the meta-model for Class II, and Random Forest for Class III. As Appendix (p11) shows, Random Forest, GBDT, and PLS-DA ranked variable importance similarly, with ESSPRI-Dryness, ESSPRI-Fatigue, ESSPRI-Pain, SGUS, subjective dry eyes, and anti-SSA consistently being the top predictors. LASSO-OR ranked variables slightly differently, highlighting the importance of hypergammaglobulinemia, cytopenia, and anti-mitochondrial antibodies. Appendix (p12) demonstrates the separation of the LCA-defined patient classes by the PLS-DA model.

The optimal sparsity parameter of the graph model ($\text{α}$) at 0.05 maximized the overall prediction accuracy (0.082) of the ordinal regression model across 10-fold CV. The variables of the Markov Blanket of the classification variable represent the minimal subset that contains the maximum information of the outcome. Figure 2 schematically demonstrates the predicted direct relationship between the 33 clinical and laboratory variables and LCA-derived patient classes. Eight variables were directly connected to LCA classes, including ESSPRI-Fatigue, ESSPRI-Dryness, subjective dry mouth, articular symptoms, anti-SSA antibody, hypergammaglobulinemia, cytopenia, and SGUS. Based on clinical expertise, we removed subjective dry mouth to reduce redundancy with ESSPRI-Dryness. The final 3-level weighted scoring system FSS, with points totaling 24, was constructed on the remaining seven items (Table 2) and achieved the accuracy of 0.806 in predicting the initial LCA-defined patient classes. Coefficients from the ordinal regression model were used to generate variable weights. Appendix (p20–21) illustrated how to apply FSS in clinical practice, using real patient cases in our cohort. A more in-depth discussion of these cases is provided in the Appendix (p4-p5).

Figure 2. Learned causal graph demonstrating the relationship between the 33 clinical and laboratory variables and LCA-derived patient classes.

Variables in yellow represent variables in the Markov Blanket, which retained essential information for inferring patient classes. Four different causal relationships were identified, including directional causal relationship: A is the cause of B; bidirectional causal relationship: unobserved latent variables caused both; conditional causal relationship: B is not a cause of A; conditional association: inconclusive causal relationship. Note that besides the edges represented by a direct arrow (AB), all other edges do not exclude the possibility of a latent confounder. AIHA: autoimmune hemolytic anemia; ANA: antinuclear antibody; eSjA: Early Sjögren’s disease autoantibody; MSGBx: minor salivary gland biopsy; RF: rheumatoid factor; RP/SG swelling: recurrent parotitis/salivary gland swelling; SGUS: salivary gland ultrasound; USFR: unstimulated salivary flow rate. This graph was generated by Cytoscape.

Table 2.

Proposed Florida Scoring System (FSS)^* for childhood Sjögren’s disease using weighted points between the minimum of 0 and maximum of 24.

Item	Weighted Points†
Subjective domain‡. Each ESSPRI score will be multiplied by the weighted points and added together. The final sum will be divided by 10.
ESSPRI-Dryness	4
ESSPRI-Fatigue	3
Objective domain. Each present item below will be multiplied by the weighted points and added together.
Cytopenia§	5
Hypergammaglobulinemia**	4.5
Anti-SSA	3
ESSDAI articular domain††	2.5
SGUS	2
Final Classification	Summation of subjective and objective scores
Class I (DDPT, Dryness Dominant with Positive Tests)	FSS score >11
Class II (HSNT: High Symptoms with Negative Tests)	6 < FSS score ≤ 11
Class III (LSNT: Low Symptoms with Negative Tests)	FSS score ≤ 6

^*FSS is applicable to any patients with clinical suspicion of childhood Sjögren’s disease. Suspected signs and symptoms include but are not limited to:

• Recurrent parotitis

• Dryness in the eyes, mouth, and/or skin with or without complications

• Positive anti-SSA/SSB and/or ANA without a definite autoimmune diagnosis

• Hypergammaglobulinemia

• Unexplained abnormal complete blood count, including neutropenia, lymphopenia, and/or anemia with or without autoimmune hemolytic anemia

• Persistently high blood amylase

• Recurrent oral, esophageal, and/or vaginal candidiasis

• Unexplained constitutional symptoms, such as fever, fatigue, and/or weight loss

• Unexplained neurologic conditions, such as aseptic meningitis, autoimmune encephalitis, dysautonomia including postural orthostatic tachycardia syndrome, gastroparesis, and/or syncope episodes

• Unexplained pain in multiple joints and/or arthritis

• Unexplained exocrine pancreatic insufficiency

• Unexplained glomerulonephritis and/or tubulointerstitial nephritis

• Unexplained chronic lung diseases

• Unexplained symptoms in primary immunodeficiency

• Unexplained immune-mediated skin conditions, such as chronic urticaria, subacute cutaneous lupus and/or vasculitis

^†Points for weighting were regression coefficients from the ordinal regression model.

^‡In the case of children who cannot report ESSPRI scores, parents can report subjective dry mouth and fatigue as “yes” or “no”. If “yes”, dry mouth will be weighted by “4” points and fatigue by “3” points instead of ESSPRI scores.

^§Cytopenia: defined as the reduction of one or more mature blood cell types (e.g., neutropenia, lymphopenia, anemia, or thrombocytopenia) in the peripheral blood.

^**Hypergammaglobulinemia: above the reference range by age

^††The ESSDAI Articular domain is a binary variable that does not require the actual ESSDAI articular score. Symptoms of arthritis or arthralgias with at least 30 min morning stiffness or synovitis will be counted as presence. Three UF patient cases with FSS applications are presented in Table S5 as examples.

Discussion

Due to the lack of childhood Sjögren’s disease specific diagnostic criteria, most providers (77%) used clinical judgment guided by, but not strictly adhering to, the 2016 ACR/EULAR adult criteria, and the vast majority (86%−99%) indicated that serologic tests were used more frequently than other diagnostic tools (e.g., SGUS, MSGBx).²⁴ However, recent multi-center retrospective studies highlighted the limitations of applying these criteria to children, in pointing out the lack of children-specific manifestations, such as recurrent parotitis, arthralgia, and neurological symptoms.^2,8,25 Unprecedentedly in the field, this data-driven study, with internally validated machine learning models applied to the only prospective childhood Sjögren’s disease cohort in the U.S.at present, proposed a pediatrician-friendly system that aims to assist clinical reasoning, disease monitoring, and follow-up plan implementation, among pediatric patients suspected of Sjögren’s. It also has important implications for future trial design, biomedical research, and therapeutics development.

We first built and validated an LCA model that identified three patient classes with distinct clinical and laboratory features. These patients are likely to differ in the disease risk, pathogenesis, disease activity, prognosis, and response to therapies. Symptom–based cluster analyses on large cohorts and clinical trials revealed adult subtypes with distinct pathobiological endotypes and responses to immunomodulatory treatments, highlighting the discordance between symptom burden, treatment, and laboratory abnormalities.^21,22 Consistently, our LCA model differentiated highly symptomatic suspected pediatric Sjögren’s patients (Class II, HSNT) from those who were only mildly symptomatic (Class III, LSNT), despite similarly negative laboratory and diagnostic profiles. This further calls attention to highly symptomatic patients undetected by the adult criteria, due to negative serological and diagnostic findings.

Previous research reported RP as the most prevalent (36%−91%) exocrine feature of childhood Sjögren’s disease.^2,8,25,26 The highest prevalence of RP in our cohort was found in Class I (56%), where 90.9% patients had a positive SGUS, implying that parotid glands might be primarily affected. Among three classes, Class I patients had the most similar presentations to pediatric Sjögren’s patients reported by a previous study,^4,26 where a high rate of RP but a lower prevalence of sicca symptoms (compared to adult patients) were reported. Class I patients in our cohort had prevalence of sicca symptoms at 56% and ESSPRI-Dryness median at 5, suggesting a moderate level of dryness, although dryness is the most dominant symptom in this group (compare to pain and fatigue). Class I also had the highest percentage (70.3%) of patients fulfilling the 2016 ACR/EULAR criteria, further highlighting its similarity to the typical presentation of adult Sjögren’s disease and possibly shared pathophysiology.

Consistent with previous findings,^2,8,25 neurological (83.9%) seemed to be most common among extraglandular manifestations, followed by articular features (80.2%), gastrointestinal (59.4%), pulmonary (53.9%), and muscular (44.2%) presentations. Class II suffered a significantly higher proportion of articular (98.0%), neurological (98.0%), gastrointestinal (82.7%), and muscular (63.0%) symptoms, suggesting possible systemic and multi-organ involvements. Among three classes, Class II also had the highest subjective dryness (sicca and ESSPRI-Dryness) with the lowest prevalence of RP (28.6%). Due to the high ESSPRI scores but low rates of positive laboratory and diagnostic findings, we are currently following up these patients to monitor whether they will progress to a full-blown Sjögren’s disease (by the adult criteria), develop another autoimmune condition, or resolve with or without medications.

Interestingly, the portions of Class II and Class III patients who fulfilled the 2016 ACR/EULAR criteria are similar (18.4% versus 14.1%), because of their similarly negative laboratory and diagnostic features. Patients in these two classes also had presentations that diverged from typical manifestations of childhood Sjögren’s disease reported by previous studies.^2,8,25,26 While it is likely that both Class II and Class III patients actually suffered from another condition, it is equally important to notice that typical serological findings (e.g., anti-SSA, anti-SSB) can be a late feature of childhood Sjögren’s disease, instead of initial presentations.²⁷ As such, the 2016 ACR/EULAR criteria may fail to capture those cases with negative laboratory features, due to its exclusive focus on objective measures, and alternative stratification strategies are therefore needed for a more careful and comprehensive evaluation.

In clinical practice, machine learning models are increasingly used to assist decision-making. Here, we tested five different models that are commonly used in clinical and electronic health record research, all of which had high accuracy (>80%) in predicting patient classes, supporting their application in the clinical setting. Variable importance ranked by different machine learning models highlighted the importance of a combination of objective (e.g., SGUS and anti-SSA) and subjective (e.g., ESSPRI) criteria, when subjective measures such as subjective dryness were previously considered infrequent in children and less important for evaluating childhood Sjögren’s disease.⁸ Based on our current study, when suspected children reported dryness, pediatricians should be alerted of a likelihood of Sjögren’s, after ruling out other possible causes.

Although machine learning models had high accuracy in predicting LCA classes, they often suffered from a lack of transparency, which makes them less attractive to clinicians. As such, we integrated clinical expertise and output from the causal graphical model, to construct the interpretable seven-item FSS for suspected pediatric Sjögren’s cases. These seven variables were consistently ranked by different machine learning models, as top predictors of LCA-derived patient classes, further validating their importance in patient stratification. Combining clinical, laboratory, and diagnostic measures, FSS captured a wide range of glandular and extraglandular manifestations that are not included in the 2016 ACR/EULAR criteria.

Cytopenia, hypergammaglobulinemia, and ESSPRI-Dryness had the highest weight in FSS. A recent study proposing a composite endpoints for adult Sjögren’s outcome measure validated the importance of ESSPRIs and serum IgG level in assessing treatment efficacy.²⁸ Furthermore, previous studies reported that hypergammaglobulinemia and cytopenia were among the most frequent laboratory findings of childhood Sjögren’s disease,^6,7 especially in those whose initial presentations were parotitis,^29,30 further supporting their particular role in evaluating pediatric Sjögren’s. Among all autoantibodies, only anti-SSA remains to be important in FSS, although some (e.g., ANA, RF, anti-dsDNA, anti-RNP, and anti-mitochondria) were also statistically significant across three classes. A previous study suggested that anti-SSA had similar contribution for characterizing children Sjögren’s cohort and adult Sjögren’s cohort.²⁶

Among all diagnostic tests (e.g., USFR, SGUS, MSGBx, Schirmer’s test), SGUS appears to be most important in distinguishing the FSS-defined classes, therefore offering a non-invasive option for young children. This is consistent with previous reported usefulness of SGUS in evaluating adult Sjögren’s.³¹ Another study reported significantly higher SGUS Hocevar score in pediatric Sjögren’s patients compared to adults, further supporting its application in the pediatric setting.²⁶ SGUS on our cohort has a sensitivity of 0.909 and specificity of 0.742 in differentiating Class I from the other two classes. Its performance in differentiating Class II from Class III was less than optimal, but those two classes did not differ significantly in all diagnostic tests. When SGUS is not available in the setting, parotid biopsy, with comparable diagnostic potential with that of MSGBx,³² can be considered as an alternative option. Considering studies of SGUS in similar pediatric conditions (e.g., sarcoidosis, IgG4-related disease) are unavailable, the sensitivity and specificity of SGUS in differentiating childhood Sjögren’s disease from those conditions need to be further investigated.

FSS score >11 indicates Class I (DDPT), which requires at least one positive serological or diagnostic findings. This class also has the highest overlap (70%) with patients who fulfilled the 2016 ACR/EULAR criteria, which may suggest they have higher risk of Sjögren’s in general, compared to the patients in the other two classes. 6 < FSS score ≤ 11 indicates Class II (HSNT), which can be achieved by only having subjective symptoms (i.e., ESSPRI-Dryness, ESSPRI-Fatigue). Due to the prominence of symptoms, periodical follow-up to monitor disease progression are highly recommended for this group. A score ≤ 6, which requires less than three positive findings, indicates Class III (LSNT). Due to the low prevalence of typical and children specific Sjögren’s symptoms as well as low rates of positive laboratory and diagnostic findings, a less stringent follow-up plan can be implemented. Combining the data-driven analysis and our past 5-year experience on this rare condition, we summarized clinical and laboratory characteristics, suggested FSS score cut-offs, and recommended potential follow-up plans for different patient classes in Table 3.

Table 3.

Summary characteristics and follow-up recommendations for patients stratified by the LCA (based on 33 variables) and FSS (based on 7 variables).

	Class I (DDPT)	Class II (HSNT)	Class III (LSNT)
Performance of using FSS to predict LCA classes: accuracy of 80.6%
FSS score cut-offs	FSS score >11	6 < FSS score ≤ 11	FSS score ≤ 6
Patients fulfilling the 2016 ACR/EULAR criteria (%)
LCA	19/27 (70%)	18/98 (18%)	13/92 (14%)
FSS	14/20 (70 %)	23/98 (23%)	13/99 (13%)
General characteristics based on LCA	Prominent glandular and extraglandular involvement with high prevalence of positive serology and SGUS	Prominent sicca and systemic symptoms with low prevalence of positive serology	Low prevalence of sicca and systemic symptoms with low prevalence of positive serology
Clinical and laboratory characteristics of each class based on LCA *
Laboratory features	High prevalence of anti-SSA (82%), anti-SSB (41%), low C3 (31%), low C4 (39%), hypergammaglobuline-mia (56%), cytopenia (41%).	Low prevalence of anti-SSA (8%), anti-SSB (5%), low C3 (1%), low C4 (9%), hypergammaglobuline-mia (1%), cytopenia(1%).	Low prevalence of anti-SSA (7%), anti-SSB (8%), low C3 (1%), low C4 (8%), hypergammaglobuline-mia (3%), cytopenia (1%).
Positive SGUS	91% (highest)	5% (lowest)	21%
Sicca prevalence	56% (high)	78% (highest)	13% (lowest)
Median Focus score	4 (highest)	0 (lowest)	1(low)
Median ESSPRI scores	Dryness (5) Fatigue (3) Pain (0)	Dryness (6) Fatigue (8) Pain (5)	Dryness (0) Fatigue (3) Pain (0)
ESSDAI	Highest in: Renal (51.9%) Cutaneous (11.1%)	Highest in: Articular (98%)† Neurological (98%)† Gastrointestinal (83%) Muscular (63%)	Lowest in: Muscular (26%) Articular (63%)
Miscellaneous	Highest prevalence of AIHA (19%)	Highest prevalence of HSD/EDS (77%)
Follow-up recommendation ‡	Every 1–3 months Closely monitor for potential organ failure and/or lymphoma.	Every 3–6 months to monitor potential progression.	Every 6–12 months to monitor potential progression.

^*Summarized based on Table 1. No statistical difference found in MGBx prevalence, Schirmer’s, RP/SG swelling, and USFR across three classes.

• MSGBx : Class I (71.4%), Class II (46.7%), and Class III (50.0%)

• Schirmer’s test: Class I (21.4%), Class II (18.2%), and Class III (18.5%)

• Objective dry mouth: Class I (37%), Class II (28.6%), and Class III (22%)

^†Although all three classes show relatively high rate of ESSDAI neurological and articular features, the highest rate is found in Class II (p=0.0005), as detailed in Table 1.

^‡‡General guidelines based on experts’ clinical experience with patients in these three classes, which are subject to change depending on each patient’s needs. All three classes are recommended for further ruling in/out childhood Sjögren’s disease using the 2016 ACR/EULAR criteria on an as-needed bases, until children-specific criteria are established in the field.

LCA: latent class analysis

FSS: Florida Score system

DDPT: Dryness Dominant with Positive Tests

HSNT: High Symptoms with Negative Tests

LSNT: Low Symptoms with Negative Tests

SGUS: Salivary gland ultrasound

ESSPRI: European League Against Rheumatism (EULAR) - Sjögren’s syndrome (SS) -patient-reported-index

ESSDAI: physician-reported European League Against Rheumatism (EULAR) - Sjögren’s syndrome (SS)-disease-activity-index.

AIHA: autoimmune hemolytic anemia

HSD/EDS: Hypermobility spectrum disorder/ Ehlers-Danlos syndrome

MSGBx: Minor salivary gland biopsy

RP/SG swelling: recurrent parotitis/salivary gland swelling.

In the absence of children specific criteria, the 2016 ACR/EULAR are commonly used in evaluating childhood Sjögren’s disease, despite low sensitivity. FSS bridges this gap, by calling attention to highly symptomatic patients who do not fulfill the adult criteria, but warrants timely intervention and close monitoring. Compared to other studies that proposed children specific criteria, ^4,9,10 we leveraged a data-driven approach for the first time in the field to construct FSS. Two clinician-guided patient-reported subjective measures (e.g., ESSPRI-Dryness, ESSPRI-Fatigue) that were considered as less important by other studies played critical role in differentiating Class II in our cohort, a patient class that requires close follow-up and maybe immediate intervention due to the prominence of their overall symptoms. We also ranked variable importance using various models, which signified the importance of the variables selected for FSS. Based on our clinical experiences and previous reports,^33,34 pediatric patients suspected of Sjögren’s may require hydroxychloroquine (HCQ) to control their symptoms. Sialendoscopy with triamcinolone acetonide was performed in children with RP.^35,36 Refractory cases of RP or significant extraglandular manifestations (particularly organ dysfunction/failure) were treated with immunosuppressive or immunomodulatory agents, such as azathioprine, mycophenolate mofetil, or leflunomide, with or without biologics such as B cell-targeted therapies ³³. Clinical trials on these patients will guarantee the establishment of evidence-based treatment strategies in the future.

The limitations of our study include: First, enrolled patients were mainly referred by pediatricians for the suspicion of childhood Sjögren’s disease, where an initial screening was conducted. This may incur selection bias and limit the generalizability of the study to a general healthy pediatric population. Second, only a limited number of patients were enrolled for the current study due to the rarity of childhood Sjögren’s disease, and an external validation cohort is unavailable. However, we performed internal validation at all stages of analyses. Third, due to the absence of children specific criteria for MSGBx, SGUS, USFR, Schirmer’s test, and autoantibodies, we used the current adult Sjögren’s cut-offs for definition positivity of these tests. Also due to the lack of specific guidelines differentiating primary from secondary Sjögren’s in children, we included suspected Sjögren’s with other suspected comorbidities. However, patients with confirmed diagnosis of other conditions that can fully explain their symptoms and laboratory findings, such as SLE, JIA, and juvenile parotitis, were ruled out. Fourth, it can be challenging for young children to rate the subjective symptoms for ESSPRI. We, therefore, included proxy variables reported by parents to evaluate subjective dryness and fatigue (subscript of Table 2) and unreliable values were considered as missing. Based on our clinical experience, children as young as 6 years old were able to provide ESSPRI scores following an instruction. Finally, our causal graphical model was currently only able to model linear relationships. However, its performance was not significantly lower than other classification methods, including those that model non-linear associations. Furthermore, all variables selected by the graph model to construct the FSS were ranked high in importance by all methods, while the graph model provides straightforward interpretability. Due to the cross-sectional design, predicted causal relationships and directions cannot be completely finalized until large-scale interventional studies become available.

In summary, in absence of a children specific criterion, we developed and internally validated FSS, a novel, clinically meaningful, and pediatrician-friendly patient stratification system, to address heterogeneous patient experience, improve the detection of childhood Sjögren’s disease, and assist disease monitoring. FSS will be particularly useful for educating and monitoring suspected symptomatic patients who do not meet the adult criteria. Our study will also serve as a steppingstone for future clinical trials and translational studies seeking to establish validated diagnostics and therapeutics customized to childhood Sjögren’s disease. Only future multicenter, large-scale, longitudinal clinical studies aiming at solidifying clinical guidelines and diagnostic criteria for childhood Sjögren’s disease will confirm the value of FSS.

Research in context.

Evidence before this study

We searched PubMed with the terms “Sjögren’s”, “Childhood Sjögren’s”, “diagnosis”, and “classification” up to August 2023, and included retrospective cohort studies, descriptive studies, and review articles from well-recognized experts. Sjögren’s disease has long been considered a disease of adulthood. Children are severely under-represented in Sjögren’s research, leading to poor understanding and possible under-diagnosis of this rare yet critical pediatric condition. Previously, descriptive analysis reported notable difference in the clinical presentations between childhood and adult Sjögren’s disease, proposing further challenges in recognizing, classifying, and monitoring the Sjögren’s disease in children. Studies investigating the performance of 2002 American-European Consensus Group (AECG) and the 2016 American College of Rheumatology/European League Against Rheumatism (ACR/EULAR) criteria found that the sensitivity of these criteria in diagnosing childhood Sjögren’s disease were unacceptably low, calling for an alternative classification strategy customized to this rare condition. Large symptom-based clustering analysis on adult Sjögren’s disease recognized discordance between symptom burden, laboratory features, and treatment responses, calling attention to Sjögren’s patients with prominent symptoms but relative negative laboratory and diagnostic profiles. A multidisciplinary expert team advocates the inclusion of EULAR Sjögren’s syndrome patient reported index (ESSPRI) as a composite outcome measure for assessing treatment efficacy, further emphasizing the importance of subjective measures in evaluating Sjögren’s disease.

Added value of this study

This study discovered three heterogeneous classes among suspected pediatric Sjögren’s disease cases, in a rare prospective pediatric cohort. Further, we determined essential variables through a clinically guided, data-driven approach and proposed a clearly-defined, pediatrician-friendly criteria for stratifying suspected pediatric Sjögren’s cases. By capturing a group of highly symptomatic pediatric patients who did not meet the 2016 ACR/EULAR criteria, the Florida Scoring System (FSS) addressed the low sensitivity concern of the adult criteria, resulting from its primary focus on objective measures. FSS also supports that salivary gland ultrasonography can be a useful non-invasive alternative to biopsy, when it comes to evaluating young patients with recurrent parotitis. Aligned with previous findings, FSS included subjective measures into the classification system, allowing a more comprehensive evaluation of childhood Sjögren’s disease that may relate to disease mechanism and treatment efficacy.

Implications of all available evidence

This study has vital implications for clinical practice as well as future large-scale observational cohorts, clinical trials, and biomedical research. First, the proposed FSS can be directly applied to the clinical setting to facilitate decision making, patient triage, and timely intervention, among pediatric patients suspected of Sjögren’s disease. Second, our study is the first step toward developing a consensus criterion for diagnosing childhood Sjögren’s disease. Finally, patients in different classes are likely to differ in pathogenesis, disease activity, prognosis, and response to therapies. As such, our findings can enlighten future biomedical research aiming at exploring the disease pathophysiology as well as clinical trials aiming at evaluating treatment efficacy.