Distinct clinical trajectories of gastrointestinal progression among patients with systemic sclerosis

Abstract

Objectives

Systemic sclerosis (SSc) is heterogeneous in its clinical presentation. Common manifestations cluster together, defining unique subgroups. This investigation aims to characterize gastrointestinal (GI) phenotypes and determine whether they can be distinguished by temporal progression.

Methods

We examined a well-established SSc patient cohort with a modified Medsger GI severity score measured over time to determine heterogeneity in disease progression. Growth mixture models estimated each patient's phenotype and disease severity trajectory over time. We compared the characteristics of estimated phenotypes using non-parametric statistics and linear and logistic regression to compare patient characteristics between phenotypes while adjusting for disease duration.

Results

We examined 2696 SSc patients with at least two Medsger GI scores, identifying four unique phenotypes. The most common phenotype (‘Stable’, n = 2325) had an average score of 1 that was consistent over time. Two phenotypes were progressive (‘Early Progressive’, n = 142, and ‘Late Progressive’, n = 115) with an initial average score of 1. The Early Progressive group increased initially and stabilized, and the Late Progressive group worsened slowly over time. A fourth phenotype (‘Early Severe GI’, n = 114) had an initial average Medsger GI score just below 3 with high mortality and improving GI severity over time.

Conclusions

Clinically distinct GI phenotypes exist among patients with SSc. These phenotypes are not only distinguished by GI and extra-intestinal SSc clinical complications, but they are also temporally distinct. Distinct autoantibody profiles are associated strongly with more severe GI disease.

Rheumatology key messages.

• Systemic sclerosis may manifest with heterogeneous GI phenotypes.

• Clinically and serologically distinct subsets of SSc GI disease can be distinguished by temporal course.

• Future research is needed to better predict these GI phenotypes and initiate early intervention.

Introduction

Systemic sclerosis (SSc) is an autoimmune rheumatic disease that may negatively impact the function of multiple organ systems and have a profound impact on the quality of life of affected persons. The GI tract is frequently impacted in SSc, although significant heterogeneity exists in clinical presentation, temporality and outcomes [1–5]. Furthermore, any part of the GI tract may be affected, from the oropharynx to the colon and anorectum, and the likelihood of progression can vary widely. These complexities introduce significant challenges in clinical practice when predicting the risk of progression, determining the timeline of GI progression and identifying patients who will develop severe GI complications that negatively impact mortality.

There is a rich history of statistical methods available for identifying subgroups from among a heterogeneous population, known as ‘unsupervised’ clustering [6], which has in recent years been applied to enhance data-driven classification of disease [7]. Many clustering methods are focused primarily on estimating patient subgroups based on a cross-section in time [8]. However, it is also of clinical interest whether patient progression over time is heterogeneous, for which a parallel yet distinct suite of statistical methodology has been developed [9]. These longitudinal clustering methods have been applied to aid in understanding disease progression, including analysis among scleroderma patients and in gastroenterology [10].

In the present report, we utilize the Johns Hopkins Scleroderma Center Research Registry to leverage 30 years of longitudinally collected data to identify distinct GI trajectories that examine GI progression over time. We aimed to identify distinct clusters that capture GI severity and differentiate SSc GI ‘progressors’ from ‘non-progressors’. We also aimed to identify demographic, extraintestinal clinical features, and distinct serological features associated with each cluster. This is the first study to utilize a large prospective registry of SSc patients to differentiate between clinically and temporally distinct GI subsets.

Methods

Study population

The Johns Hopkins Scleroderma Center Research Registry (JHSCRR) was established in 1990 to study the natural history of the disease among patients seen in our centre. We performed a retrospective analysis of all patients seen between January 1990 and January 2020 who had at least two visits with two modified Medsger GI severity scores captured in the database. Patient data before 2000 was compiled retrospectively, and from 2000 and later were collected prospectively for all patients in our longitudinal database if they met 2013 American College of Rheumatology (ACR)/European League Against Rheumatism (EULAR) criteria for systemic sclerosis [11], 1980 ACR criteria [12], had at least three of five features of CREST syndrome, or definite Raynaud's, abnormal nail fold capillaries, and an SSc-specific autoantibody [13]. At each visit, an assessment of the clinical involvement of SSc and a detailed medication reconciliation were recorded.

Clinical phenotyping

The Center's database also includes demographic and clinical data, including age at first clinical visit, sex, race, smoking status, disease duration from first Raynaud's or non-Raynaud's symptom, SSc skin subtype, and autoantibody status. Medsger severity scores and objective measures define specific organ involvement at baseline, and then approximately every 6 months at follow-up visits on all actively followed patients. Clinical information related to SSc patient presentation and management was collected over time, including GI symptoms specified by the modified Medsger GI score (Table 1). Mild GI disease was defined by a modified Medsger severity score of 0 (i.e. normal) or 1 [i.e. requiring gastroesophageal reflux disease (GERD) medications or abnormal small bowel series]. Moderate GI disease was characterized by a Medsger GI score of 2, which involves the use of high doses of GERD medications and/or an abnormal small bowel series. Severe GI disease was documented by a Medsger GI severity score of 3 (i.e. evidence of recurrent pseudo-obstruction or malabsorption) or a Medger GI severity score of 4 (i.e. requiring total parenteral nutrition). The presence of lung and/or cardiac involvement was defined by a score of 1 or greater on the respective Medsger severity scores (lung or heart) [6]. Severe Raynaud's was defined as a Medsger Raynaud's severity score of ≥2 (i.e. presence of digital pits, ulcers or gangrene). Myopathy was identified by the physician and dependent on the presence of at least one of the following abnormal test results at any visit: an elevation in creatinine phosphokinase, myopathic features on electromyography, MRI consistent with muscle inflammation, or abnormal muscle biopsy. Tendon friction rubs and cancer (ever/never) were also recorded. Specific measures of pulmonary function were recorded [forced vital capacity (FVC) and single-breath diffusing capacity for carbon monoxide (DLCO), measured as the absolute value, as well as the percentage predicted value, all standardized for sex and age]. Estimated right ventricular systolic pressure (RVSP) was measured by transthoracic echocardiogram and obtained as part of routine clinical screening. All analyses were performed using the maximum (for RVSP and Medsger scores) or minimum (FVC, DLCO) score estimate for any study. Data on the use of opioids, promotility agents and immunosuppressive medications (mycophenolate mofetil, methotrexate, cyclophosphamide and intravenous immunoglobulin) were collected coincidentally with other clinical variables. Information on death was collected through the Social Security Death Index, the standard practice in managing our database. The end of follow-up was determined by the last visit/patient contact date. The measurement and definition of autoantibody status is described in Supplementary Appendix 1, available at Rheumatology online. Written informed consent was obtained from all patients. The present study was approved by the Johns Hopkins Institutional Review Board.

Table 1.

Modified Medsger GI severity score

Score	Description
0	Normal
1	GERD meds OR abnormal small bowel series
2	High-dose GERD meds OR antibiotics for bacterial overgrowth
3	Malabsorption syndrome OR episodes of pseudo-obstruction
4	TPN required

GERD: gastroesophageal reflux disease; GI: gastrointestinal; meds: medications; TPN: total parenteral nutrition.

Clustering and other statistical methods

Patient data over time were compiled so that the first visit to the Scleroderma Center for each patient was set as time zero, and time was measured forward from this date of the earliest recorded Medsger GI severity score. To cluster trajectories in Medsger score over time, we used an extension of the linear mixed model having random patient-level intercepts, with a growth mixture model assuming a finite number of latent groups within the patient population specified by a distinct trajectory, using methods defined by Proust-Lima et al. [14]. This growth mixture model specified time using a B spline basis with three degrees of freedom [15], allowing for non-linear changes over time, as expected if patients’ trajectories changed direction or did not progress beyond a certain level. This model has three equally spaced knots, or points where the polynomial function of time is allowed to change and four phenotype-level effects for each knot. These phenotype-level effects are jointly distributed with a mixture of Gaussian distributions such that the expected patient Medsger GI score is a function of time and phenotype, with an approximately Gaussian error.

To determine the number of unique groups or phenotypes, we used a two-step cross-validation procedure, as defined by Fu and Perry [16]. First, for a fixed number of phenotypes, ranging from two to six, we randomly sampled half of the patients and estimated the trajectory of Medsger GI scores among each phenotype over time. Then, for each patient who did not contribute to the trajectory estimation, we sampled half of their recorded Medsger scores, used the chronological earlier half to estimate the phenotype to which they most likely belonged, and then predicted their Medsger scores for the times in the chronologically latter half. Then, the difference between the predicted and recorded Medsger scores for the second half was used to estimate the cross-validation error. This cross-validation error was then averaged across all predicted Medsger scores, and the process was repeated over 1000 random samples.

We selected the number of distinct phenotypes based on the lowest cross-validation error. Using the number of phenotypes indicated by the cross-validation, we estimated the phenotypes using the latent class linear mixed model described above among all eligible patients, yielding an estimated probability for each patient/phenotype combination. The phenotype with the highest probability was chosen as the most likely for each patient. To compare features between estimated phenotypes, we used the non-parametric Fisher’s exact test (for categorical features) or the Kruskal–Wallis test (for continuous features). To examine the mortality outcome between phenotypes, we examined the Kaplan–Meier survival over the follow-up period, with time zero being the first visit time and follow-up concluding at the last visit or time of death. We compared mortality between phenotype groups with the log-rank test. In addition to these comparisons, we used logistic and linear regression to examine the differences between phenotypes for each feature. Given its critical importance in symptom manifestation, we adjusted for disease duration from the first SSc symptom (either Raynaud’s or non-Raynaud’s).

Given that patients may differ in their GI progression based on their overall SSc disease duration before presentation in our clinic, we devised a sensitivity analysis limited to patients <5 years from the time of the first Raynaud’s phenomenon or non-Raynaud’s phenomenon symptom, whichever came first. We performed additional sensitivity analyses to determine whether our results from the original analysis were impacted significantly by mortality (i.e. if patients with high Medsger GI scores were more likely to die early and thus alter the trajectory), as well as recall (excluding patients with Medsger GI scores measured in the retrospective period before 2000).

Results

From 1990 until January 2020, 3935 distinct SSc patients were evaluated at the Scleroderma Center with at least one measured Medsger GI score, with 2696 (69%) having two or more visits with the Medsger GI score. Of these 2696 patients, the median number of visits was 6, with a mean of 9 visits and a range of 2–45 trips. On average, patient follow-up time spanned 6.7 years, with a median (interquartile range) of 4.7 (1.9–10.3) years.

These 2696 patients are described in Table 2. Most patients (82%) were female, with a mean age of 52 years and disease duration from the first symptom, either Raynaud’s or non-Raynaud’s, of 9 years (s.d. 10.7) at their first visit. Regarding SSc subtype, most patients were either limited (61%, of which 5% were sine) or diffuse (39%), and the majority of patients were White (76%). Anti-centromere (CENP) and anti-Ro52 were the most common autoantibodies in these patients at 28%, followed by anti-topoisomerase-1 (TOPO) at 22%.

Table 2.

Demographic and clinical features of 2696 patients with SSc having at least two clinical visits with measured Medsger GI scores

Characteristic	n measured	Value
Female, n (%)	2696	2224 (82)
Age at first visit, median (IQR), years	2696	52.17 (42.2, 61.8)
Follow-up time, median (IQR), years	2696	4.66 (1.9, 10.3)
Disease duration, median (IQR), years from first RP	2633	4.62 (1.6, 12.5)
Disease duration, median (IQR), years from first symptoma	2693	3.94 (1.1, 11.9)
Subtype, n (%)	2694
Limited cutaneous disease (including sine)		1652 (61)
Diffuse cutaneous disease		1042 (39)
Race (self-identified), n (%)	2696
White		2055 (76)
Black		467 (17)
Other		175 (6)
Renal crisis (ever), n (%)	2696	122 (5)
Cancer (ever), n (%)	2696	522 (19)
Medsger score, median (IQR)
Maximum RP	2696	2.0 (1.0, 3.0)
Maximum lung	2590	0.0 (0.0, 2.0)
Maximum heart	2583	2.0 (1.0, 2.0)
Maximum GI	2696	1.0 (1.0, 1.0)
Severe RP (maximum >1), n (%)b	2696	1626 (60)
Severe lung (maximum >1), n (%)b	2590	1774 (66)
Severe heart (maximum >1), n (%)b	2583	706 (26)
Severe GI (maximum >1), n (%)b	2696	1450 (54)
Telangiectasia, n (%)	2696	2537 (94)
Calcinosis, n (%)	2696	1013 (38)
Dry eye, n (%)	2696	1528 (57)
Dry mouth, n (%)	2696	1625 (60)
Myopathy (ever), n (%)d	2696	560 (21)
Maximum mRSS, median (IQR)	2655	7.0 (3.0, 18.0)
Minimum FVC, median (IQR), % predicted	2612	75.0 (58.7, 90.0)
Minimum DLCO, median (IQR), % predicted	2579	55.0 (36.8, 70.4)
Maximum RVSP, median (IQR), mmHg	2145	38.0 (31.0, 51.0)
Autoantibodies, n (%)c
Centromere (CENP)	2105	581 (28)
ThTo	2105	171 (8)
Ro52	2105	582 (28)
Ku	2105	94 (4)
PM-Scl	2105	43 (2)
Topoisomerase-1 (TOPO)	2105	467 (22)
RNA polymerase-3 (POL3)	2105	354 (17)
Fibrillarin (U3RNP)	2105	148 (7)
Nor90	2105	87 (4)
U1RNP	2101	259 (12)
RNPC3	2103	118 (6)

^aTime from first symptom at first visit, where first symptom is defined as either Raynaud’s or non-Raynaud’s.

^bSeverity measured by Medsger score.

^cAutoantibody measures based on Eurimmune assay.

^dMyopathy was defined by evidence of elevated muscle enzymes, EMG abnormalities consistent with myopathy, a positive muscle biopsy, or abnormal MRI consistent with myopathy. DLCO: diffusing capacity of the lungs for carbon monoxide; EMG: electromyography; FVC: forced vital capacity; GI: gastrointestinal; IQR: interquartile range; mRSS: modified Rodnan skin score; RP: Raynaud’s phenomenon; RVSP: right ventricular systolic pressure.

We further characterized patients by applying a hierarchical latent class linear mixed model to their Medsger GI scores over time [15]. First, we determined the number of distinct phenotypes represented among our patients [17]. Cross-validation error indicated that four phenotypes were the most appropriate choice (Supplementary Fig. S1, available at Rheumatology online). We fit the full hierarchical latent class linear mixed model with the four estimated phenotypes among the complete data. We also examined the standardized residuals from this model compared with qualities from the standard normal distribution (Supplementary Fig. S2, available at Rheumatology online).

Estimated trajectories of GI function in these four phenotypes are shown in Fig. 1. The most common phenotype (cluster C, ‘Stable’ 86%) was defined by a generally low and consistent Medsger GI score over 15 years of observation. The next most common phenotype (cluster A, ‘Early Progressive’, 5%) had an estimated trajectory that began near 1 (at enrolment), increased initially over the first 10 years of follow-up, and then remained relatively constant. Another phenotype (cluster D, ‘Late Progressive’, 4%) had an increasing Medsger GI score, but it began slowly and later progressed more rapidly. The last phenotype (cluster B, ‘Early Severe GI’, 4%) had severe disease at presentation with decreasing Medsger GI scores over the 15 years. Our estimated phenotypes were not sensitive to diagnosis time, mortality or recall, as the estimated trajectories appeared similar in sensitivity analyses, described in supplementary material and in Supplementary Fig. S3, available at Rheumatology online. A complete description of the four estimated phenotypes is shown in Table 3.

Figure 1.

Estimated Medsger score trajectories among 15 years of follow-up SSc in 2696 scleroderma patients with at least two visits, assuming there are four distinct phenotypes. Two sensitivity analyses are included: (i) among 1399 patients with GI scores in the first 5 years of SSc, and (ii) among all 2696 patients carrying forward the last GI scores among those with mortality. GI: gastrointestinal

Table 3.

Description of four phenotypes among 2696 scleroderma patients with at least two measured Medsger GI scores, mean or percent by phenotype

	Phenotype
Characteristic	Cluster A	Cluster B	Cluster C	Cluster D	P
	(n = 142)	(n = 114)	(n = 2325)	(n = 115)
Female, n (%)	112 (79)	85 (75)	1926 (83)	101 (88)	0.036
Age at first visit, median, years	58.5	58.5	52.2	51.1	0.005
Follow-up time, median, years	4.1	4.1	4.3	11.1	<0.001
Disease duration, median, years from first symptom	6.4	6.4	4.8	3.3	<0.001
Disease duration, median, years from first RP	6.3	6.3	4.0	3.2	<0.001
Subtype, n (%)
Limited cutaneous disease (including sine)	26 (39)	64 (56)	1464 (63)	70 (61)	<0.001
Diffuse cutaneous disease	86 (61)	50 (44)	861 (37)	45 (39)	<0.001
Race, n (%)
White	105 (74)	88 (77)	1777 (76)	85 (74)	0.814
Black	30 (21)	22 (19)	393 (17)	22 (19)	0.470
Other	7 (5)	4 (4)	156 (7)	8 (7)	0.534
Mortality (yes), n (%)	66 (46)	65 (57)	765 (33)	47 (41)	<0.001
Renal crisis (ever), n (%)	10 (7)	6 (5)	101 (4)	5 (4)	0.431
Cancer (ever), n (%)	25 (18)	27 (24)	442 (19)	28 (24)	0.388
Myopathy (ever), n (%)	51 (36)	30 (26)	453 (19)	26 (23)	<0.001
Maximum RP, median, Medsger score	2.0	2.0	2.0	2.0	<0.001
High maximum RP (Medsger score >2), n (%)	60 (42)	29 (25)	642 (28)	54 (47)	<0.001
Maximum lung, median, Medsger score	2.0	2.0	2.0	2.0	<0.001
High maximum lung (Medsger score >2), n (%)	65 (47)	50 (46)	836 (38)	54 (47)	<0.001
Maximum heart, median, Medsger score	0.0	0.0	0.0	0.0	<0.001
High maximum heart (Medsger score >2), n (%)	35 (26)	29 (26)	394 (18)	35 (30)	<0.001
Maximum GI, median, Medsger score	3.0	3.0	1.0	2.0	<0.001
High maximum GI (Medsger score >2), n (%)	80 (56)	89 (78)	99 (4)	32 (28)	<0.001
Telangiectasia, n (%)	135 (95)	109 (96)	2178 (94)	115 (100)	0.026
Calcinosis, n (%)	54 (38)	47 (41)	837 (36)	75 (65)	<0.001
Dry eye, n (%)	84 (59)	71 (62)	1290 (55)	83 (72)	0.002
Dry mouth, n (%)	88 (62)	80 (70)	1370 (59)	87 (76)	<0.001
Maximum mRSS, median	9.0	9.0	6.0	11.0	<0.001
Minimum FVC, median, % predicted	71.8	71.8	76.0	71.0	<0.001
Minimum DLCO, median, % predicted	53.0	53.0	55.3	46.0	0.002
Maximum RVSP, median, mmHg	40.0	40.0	38.0	40.0	0.046
Autoantibodies, n (%)
Centromere (CENP)	26 (21)	23 (26)	494 (28)	38 (35)	0.136
ThTo	10 (8)	7 (8)	149 (8)	5 (5)	0.599
Ro52	35 (28)	32 (36)	483 (27)	32 (29)	0.356
Ku	6 (5)	3 (3)	81 (5)	4 (4)	0.955
PM-Scl	1 (1)	0 (0)	39 (2)	3 (3)	0.434
Topoisomerase-1 (TOPO)	27 (22)	10 (11)	398 (22)	32 (29)	0.018
RNA-polymerase 3 (POL3)	21 (17)	8 (9)	309 (17)	16 (15)	0.177
Fibrillarin (U3RNP)	17 (14)	9 (10)	114 (6)	8 (7)	0.018
PDGFR	0 (0)	0 (0)	1 (0)	0 (0)	1.000
NOR90	6 (5)	6 (7)	71 (4)	4 (4)	0.544
U1RNP	18 (15)	16 (18)	215 (12)	10 (9)	0.233
RNPC3	10%	12%	5%	4%	0.004

Cluster A: Early Progressive; cluster B: Early Severe GI; cluster C: Stable; cluster D: late progressive. Significance for a difference between any two phenotypes, determined by Fisher’s exact test for categorical features (e.g. gender) or the non-parametric Kruskal–Wallis test for continuous features (e.g. disease duration). DLCO: diffusing capacity of the lungs for carbon monoxide; FVC: forced vital capacity; GI: gastrointestinal; mRSS: modified Rodnan skin score; RP: Raynaud’s phenomenon; RVSP: right ventricular systolic pressure.

The Early Severe GI patients were older, more likely to be male, had more sicca symptoms and had the second highest prevalence of myositis across the groups (Table 3). When examining autoantibody profiles across the four phenotypes, anti-fibrillarin (U3RNP) antibodies were more common among the Early Severe GI group patients compared with those in the Late Progressive group and the Stable group. A description of the overlap of autoantibodies between phenotypes adjusted for multiple comparisons using the Benjamini–Hochberg method is in Supplementary Table S1, available at Rheumatology online.

Patients in the Early Progressive group were more likely to have diffuse cutaneous fibrosis. They were the most likely to have myositis than patients in the other groups and had higher Medsger lung severity scores (>2) (Table 3). Importantly, these patients had the second highest Raynaud’s severity scores. The Early Progressive group was distinguished from the Late Progressive and Stable groups by having a high prevalence of anti-fibrillarin (U3RNP) antibodies.

The Late Progressive group had the highest proportion of women at 88%, the youngest at the first visit, and the shortest average disease duration. They also had the most severe vascular findings across the groups, including the lowest minimum DLCO (46.0, s.d. 23.4, Table 3), which is often considered a marker of small vessel disease in the lungs and a risk factor for pulmonary hypertension (PH) [18–20]. Furthermore, the Late Progressive group had the highest prevalence of cardiac involvement (30% with a Medsger heart score >2) and telangiectasia. This group also had the highest frequency of TOPO across the groups and was least likely to have anti-U1RNP and anti-RNPC3 antibodies.

The Stable group is the most common phenotype among the patient cohort and generally the least severe. These patients most often had limited cutaneous disease and a relatively low frequency of myopathy but had more severe Raynaud’s phenomenon. Patients in this cluster also had the lowest average Medsger scores for respiratory and cardiovascular health and the lowest rates of calcinosis and sicca symptoms.

The disease characteristics and demographics among the four estimated phenotypes were also compared after adjusting for disease duration from the first symptom (Table 4). The phenotypes are generally slightly more similar after adjusting for disease duration, particularly in age and telangiectasia, calcinosis, and sicca prevalence. However, many features remained significantly different across phenotypes even after accounting for disease duration, including race, maximum mRSS score, myopathy, minimum FVC, minimum DLCO and autoantibody types [TOPO, fibrillarin (U3RNP) and RNPC3].

Table 4.

Estimated differences and 95% confidence intervals between phenotypes relative to cluster C (‘Stable’ group) and adjusted for disease duration from time of first symptom (n = 2696), shown as odds ratios for categorical factors and average differences for continuous factors

	Phenotype (relative to C)
Characteristic	Cluster A	Cluster B	Cluster D	P (difference overall) c
Femalea	0.58 (0.35, 0.94)	0.51 (0.31, 0.85)	0.99 (0.52, 1.91)	0.017
Age at first visit (years)b	0.33 (−2.36, 3.02)	3.59 (0.74, 6.44)	0.52 (−2.67, 3.70)	0.105
Follow-up time (years)	−2.02 (−3.27, −0.78)	−1.89 (−3.20, −0.57)	−0.76 (−2.23, 0.71)	0.001
Subtype a
Limited cutaneous disease (including sine)	0.32 (0.20, 0.51)	0.55 (0.34, 0.89)	1.28 (0.74, 2.19)	<0.001
Diffuse cutaneous disease	3.13 (1.97, 4.95)	1.81 (1.13, 2.90)	0.78 (0.46, 1.34)	<0.001
Race a
White	1.00 (0.62, 1.62)	0.76 (0.46, 1.24)	1.03 (0.58, 1.84)	0.724
Black	1.16 (0.69, 1.96)	1.69 (1.00, 2.85)	0.92 (0.48, 1.78)	0.241
Other	0.66 (0.27, 1.64)	0.56 (0.19, 1.64)	1.06 (0.41, 2.75)	0.547
Mortalitya	2.82 (1.86, 4.27)	3.80 (2.44, 5.93)	2.25 (1.37, 3.70)	<0.001
Renal crisisa	1.68 (0.72, 3.93)	2.06 (0.85, 4.98)	1.14 (0.35, 3.65)	0.352
Cancera	0.78 (0.45, 1.35)	1.36 (0.81, 2.26)	1.00 (0.54, 1.85)	0.483
Myopathy (ever)a	2.50 (1.60, 3.91)	1.66 (1.00, 2.74)	0.72 (0.38, 1.38)	<0.001
Maximum RPb	0.25 (0.06, 0.44)	−0.07 (−0.28, 0.13)	0.16 (−0.07, 0.39)	0.029
Maximum lungb	0.34 (0.07, 0.61)	0.43 (0.14, 0.72)	0.11 (−0.21, 0.42)	0.004
Maximum heartb	0.37 (0.04, 0.69)	0.42 (0.08, 0.75)	0.45 (0.07, 0.82)	0.006
Maximum GIb	1.81 (1.67, 1.95)	2.22 (2.07, 2.36)	0.43 (0.27, 0.59)	<0.001
Telangiectasiaa	0.99 (0.94, 1.04)	1.01 (0.96, 1.06)	1.01 (0.95, 1.07)	0.953
Calcinosisa	0.78 (0.50, 1.21)	1.11 (0.71, 1.73)	1.30 (0.79, 2.14)	0.421
Dry eyea	0.98 (0.65, 1.48)	1.35 (0.86, 2.12)	0.91 (0.56, 1.48)	0.531
Dry moutha	1.04 (0.68, 1.58)	1.98 (1.22, 3.22)	1.06 (0.64, 1.74)	0.041
Maximum mRSSb	6.72 (4.38, 9.05)	4.42 (1.97, 6.86)	1.68 (−1.03, 4.39)	<0.001
Minimum FVC (% predicted)b	−7.67 (−12.16, −3.18)	−6.73 (−11.46, −2.00)	−1.74 (−6.99, 3.52)	0.001
Minimum DLCO (% predicted)b	−5.82 (−10.77, −0.87)	−5.69 (−10.89, −0.50)	−3.57 (−9.33, 2.19)	0.026
Maximum RVSP (mmHg)b	2.29 (−2.46, 7.03)	−0.24 (−5.11, 4.63)	4.74 (−0.47, 9.95)	0.258
Autoantibodiesa
Centromere (CENP)	0.79 (0.45, 1.38)	0.74 (0.40, 1.35)	1.84 (1.05, 3.22)	0.061
ThTo	0.74 (0.32, 1.72)	0.83 (0.33, 2.06)	0.41 (0.14, 1.25)	0.378
Ro52	0.91 (0.55, 1.51)	1.32 (0.78, 2.24)	1.13 (0.65, 1.96)	0.715
Ku	0.96 (0.34, 2.72)	0.79 (0.22, 2.84)	0.79 (0.22, 2.81)	0.973
PM-Scl	0.09 (0.00, 1.86)	0.03 (0.00, 5.83)	0.99 (0.19, 5.03)	0.053
Topoisomerase-1 (TOPO)	0.76 (0.44, 1.32)	0.41 (0.20, 0.87)	1.20 (0.67, 2.14)	0.040
RNA polymerase-3 (POL3)	0.85 (0.47, 1.51)	0.43 (0.18, 1.01)	0.69 (0.34, 1.41)	0.181
Fibrillarin (U3RNP)	3.06 (1.58, 5.93)	3.06 (1.40, 6.69)	1.30 (0.49, 3.46)	0.002
NOR90	1.17 (0.40, 3.46)	2.21 (0.84, 5.84)	1.12 (0.32, 3.91)	0.522
U1RNP	1.06 (0.55, 2.05)	1.82 (0.96, 3.45)	0.65 (0.28, 1.51)	0.166
RNPC3	2.76 (1.32, 5.76)	2.20 (0.92, 5.26)	0.76 (0.22, 2.58)	0.026

Cluster A: Early Progressive; cluster B: Early Severe GI; cluster C: Stable; cluster D: Late Progressive.

^aEstimated differences with phenotype C shown as odds ratios, adjusted for disease duration, estimated with logistic regression.

^bEstimated differences with phenotype C shown as mean differences, adjusted for disease duration, estimated with linear regression.

^cSignificance of any difference between phenotypes determined with a likelihood ratio test, comparing the regression model with covariates for phenotype membership to the regression model that does not include any covariate for phenotype. DLCO: diffusing capacity of the lungs for carbon monoxide; FVC: forced vital capacity; GI: gastrointestinal; mRSS: modified Rodnan skin score; RP: Raynaud’s phenomenon; RVSP: right ventricular systolic pressure.

Medications taken during follow-up by estimated phenotypes are shown in Supplementary Table S2, available at Rheumatology online, grouped by indication. Medication use was generally the lowest among the Stable group (34% vs 44%, 44% and 60% for clusters A, B and D, respectively) for medications such as antibiotics and opioids. It was also low compared with other phenotypes for treating acid reflux disease (81% vs 85%, 89% and 98%). The Late Progressive group had high medication use generally, particularly for the treatment of reflux (98%).

The Kaplan–Meier estimated survival of all 2696 patients is shown in Fig. 2 for up to 30 years of follow-up. Survival was highly variable between phenotypes (P < 0.001). The highest mortality rate was observed among the Early Severe GI group at 57%, occurring early during follow-up. The Early Progressive group also had a high mortality rate early on during follow-up, in contrast to patients in the Stable group, who had a relatively consistent mortality rate over the period. The most favorable mortality profile was among patients in the Late Progressive group.

Figure 2.

Comparison of mortality between different estimated phenotypes for 2969 scleroderma patients. Survival was compared between phenotypes using the log rank test

Discussion

This is the first study to evaluate patterns in longitudinal GI trajectories among patients with SSc to identify high-risk GI groups through an unbiased trajectory modelling approach. We utilized data from one of the largest clinical datasets available in SSc housing well-phenotyped patients. We then applied sophisticated longitudinal clustering methods and identified four distinct and clinically meaningful GI trajectories in patients with SSc, each distinguished by relative degrees of GI severity, time course, extra-intestinal clinical features and autoantibody profiles. This study lays an important foundation for further refining clinical GI subsets in SSc and pursuing translational studies among more homogeneous subgroups to shed light on disease mechanisms.

Our unbiased GI trajectory analysis identified four distinct SSc GI clinical subgroups. The Early Severe GI subgroup was distinct from others in that patients had the highest (worst) GI severity scores and were more likely to be male than other subgroups. These patients were also older in age, more likely to die, had more sicca symptoms, and were the second highest group with myositis. These risk factors for severe GI disease have been identified in other studies that validate this cluster [3, 5, 21, 22]. Interestingly, cardiopulmonary disease was not common or severe in these patients, and anti-TOPO antibodies, which are associated with more severe and progressive interstitial lung disease, were also uncommon in this group. The most common antibodies among patients in this severe GI group were anti-RNPC3, anti-U1RNP and anti-Ro52. This group also had the second-highest proportion of anti-fibrillarin (U3RNP) antibodies. These clinical and serological associations with severe GI disease are validated by other smaller GI-focused studies in the published literature [20, 23, 24].

The Early Progressive group was the second most severe group with the second highest GI scores, though it had the shortest SSc disease duration among the four groups. Early Progressive patients, however, were distinct from the Early Severe GI group as they were more likely to have significant fibrosis, pulmonary hypertension and severe Raynaud’s. This group had the second-highest proportion of males and the second-highest mortality. Given that the Early Progressive group had the highest proportion of anti-fibrillarin (U3RNP) antibodies (though not statistically significant), the finding of significantly higher RVSPs (i.e. suspected PH) was not surprising as PH is more common among anti-fibrillarin (U3RNP) positive patients [25]. Furthermore, given this group's high prevalence of significant interstitial lung disease (ILD), some of the PH was likely secondary to ILD. The association between short disease duration and severe GI disease has been reported, particularly in a subset of patients with recurrent pseudo-obstruction [1, 4]. It was also striking that anti-RNPC3 and anti-U1RNP antibodies were the second most prominent in this group with the second highest GI severity (only less prominent than the Early Severe GI group). This strong association between anti-RNPC3 and anti-U1RNP antibodies and GI severity in our study again suggests that both anti-RNPC3 and U1RNP antibodies are very important specificities in patients with more severe GI disease in SSc and corroborates such evidence from a much smaller study in the published literature [22].

In contrast to the prior two groups, the Late Progressive group had significantly more women and significantly worse Medsger Raynaud’s severity scores. Furthermore, all patients in this group had telangiectasia, and calcinosis was most prevalent (65% vs 41%). This is interesting as calcinosis is often considered a complication of vasculopathy in SSc [26]. This group also had the lowest DLCO, which has historically been attributed to progressive microvascular disease, and the highest RVSP, which is associated with pulmonary hypertension. Finally, this group had the highest proportion of anti-CENP and anti-TOPO-positive patients across the groups, both of which are associated with more significant vascular complications in SSc [27–29]. Similar to the Stable group, patients in this group had moderate GI disease over time, though Medsger GI scores in the Late Progressive group were increasing more than in the Stable group.

Patients in the Stable group had the mildest GI disease and the lowest mortality overall. Patients in this group were more likely to be women, were younger in age, and had low heart and lung severity scores. Among patients in this group, more patients had limited cutaneous disease. The most commonly observed autoantibodies among patients in the Stable group were antibodies to centromere, Ro52 and TOPO. Importantly, most patients with anti-TOPO antibodies in this cluster had limited cutaneous disease, suggesting that anti-TOPO-positive patients with limited cutaneous disease may be less likely to have GI progression.

Our study has many strengths. First, we utilized a large, well-phenotyped longitudinal cohort to identify patients with distinct clinical GI trajectories objectively. In doing so, we also identified the extraintestinal features of these patients. This was coupled with significant expertise in statistical methods concerning the clustering analyses. Finally, we identified specific serological profiles that may help to predict these trajectories, as serologies are often present years before disease onset [30–32]. Our study also has limitations. First, we have yet to fully assess the impact of pharmacological treatment on each of the GI clusters. However, as no medications are known to impact the progression of GI disease in scleroderma, we do not suspect these data would significantly impact our clusters. Second, we utilized the modified Medsger GI severity score, the GI severity measure collected in our database. While the Medsger GI severity score does not incorporate objective measures of GI motility, we have completed several studies that help us to understand better the types of dysmotility present among patients in each of these severity groups [33, 34]. This modified Medsger GI severity score may also impact the comparison of GI clusters to other cohorts and studies. Our patient cohort is almost exclusively American, so it is not necessarily generalizable to scleroderma in other populations. The classification criteria for SSc have changed over time, and we did not systematically capture the Very Early Diagnosis of Systemic Sclerosis (VEDOSS) in our database; therefore, the absence of this classification could potentially lead to missing patients with very early SSc, as this is a more recent definition. It is difficult to rule out unmeasured confounding in the associations between estimated phenotypes and features of SSc disease. Finally, we used indirect measures of ILD, PH and vascular disease in characterizing SSc patients.

Here, we estimated a phenotype for each patient to estimate the cross-validation error and determine the number of unique phenotypes. The average squared error for the selected four phenotypes was around 0.45 (Supplementary Fig. S1, available at Rheumatology online), corresponding to within 0.7 points in the modified Medsger GI score. Whether or not this level of prediction error is acceptable would depend on the specific application of these phenotypes. In practice, some patients are likely to have more precisely estimated phenotypes than others (for example, patients with more visits). Potential applications using these phenotypes are variable, including counselling patients regarding their likely progression and management and identifying patients likely to progress for novel biomarker detection. The specific applications are the subject of future research.

Conclusions

Our research strongly supports the idea that clinically distinct GI phenotypes exist among patients with SSc and many patients have mild or stable GI disease. Importantly, these phenotypes are not only distinguished by GI and extra-intestinal SSc clinical complications, but they are also temporally distinct. The identification of temporally based clusters is helpful in identifying patients who may benefit most from a GI-focused clinical trial early vs late in their SSc. Also, it helps to define phases within SSc where patients are at the highest risk for progressive GI disease. While we examined factors across many years to estimate cluster membership for each patient, for clinical utility, cluster membership would need to be predicted early in the course of the disease. Identifying homogeneous subgroups would also create a framework from which the study of disease mechanisms across similar patients is more feasible. Finally, identifying distinct autoantibody profiles strongly associated with more severe GI disease will allow the integration of predictive biomarkers in clinical care and help investigators enrich the subgroups of patients most likely to progress in GI-related clinical trials.

Supplementary material

Supplementary material is available at Rheumatology online.

Data availability

Data are available for research purposes upon request.

Funding

The authors are grateful to the support from NIH/NIAMS 1R01AR081382; Rheumatology Research Foundation to Z.M. as well as P30 AR070254, K24 AR080217, the Donald B. and Dorothy L. Stabler Foundation, the Jerome L. Greene Foundation, the Chresanthe Staurulakis Memorial Discovery Fund, the Martha McCrory Professorship, the Manugian Family Scholar, the Nancy and Joachim Bechtle Precision Medicine Fund for Scleroderma, and the Johns Hopkins inHealth initiative.

Disclosure statement: Z.M. is a consultant for Boehringer Ingelheim.