Background

Systemic sclerosis (SSc) is a chronic autoimmune disease characterised by extensive fibrosis of the skin, internal organs and vasculopathy [1]. Interstitial lung disease (ILD) is a common manifestation of SSc, in which parenchymal involvement can lead to pulmonary fibrosis and declining pulmonary function [2, 3]. SSc-associated ILD (SSc-ILD) is associated with reduced survival and is the leading cause of death in patients with SSc [4].

Forced vital capacity (FVC) is currently supported as an endpoint for randomised controlled trials and is the primary outcome measure in many of them [5, 6]. An FVC absolute decline of at least 10% is an established measure of progression in more progressive ILDs like idiopathic pulmonary fibrosis (IPF) [7], but lower levels of progression in SSc-ILD may better reflect the disease course. For instance, an FVC absolute decline of 5–9% with a diffusing capacity of the lung for carbon monoxide (DLCO) absolute decline of ≥ 15% has been considered optimal for trial purposes [8], and an FVC decline or improvement of approximately 3% was considered a minimal clinically important difference based on data from Scleroderma Lung Study I and II [6].

Although FVC is important, patient-reported outcome (PRO) measures have recently received growing recognition as they can provide valuable insights about the impact of a disease on patients and drive improvements in care [9]. Studies have shown that cough, dyspnoea, changes in physical function and fatigue are the lung-related symptoms of most concern to patients with connective tissue disease (CTD)-associated ILD, including SSc [10, 11]. Thus, PRO measures are usually assessed in randomised clinical trials in SSc-ILD as secondary endpoints [12].

SENSCIS® was a large, phase III, randomised, placebo-controlled, parallel trial of nintedanib, a multi-target tyrosine kinase inhibitor with anti-inflammatory and antifibrotic properties, in patients with SSc-ILD [13]. SENSCIS® demonstrated a statistically significant 44% reduction in ILD progression in patients with SSc-ILD receiving nintedanib compared with those who received placebo. For PRO measures, very small changes were observed over 52 weeks, with no differences between nintedanib and placebo. Post hoc analysis showed that patients with more impaired lung function (FVC < 70% predicted at baseline) had worse health-related quality of life (HRQoL; as indicated by higher St. George’s Respiratory Questionnaire values) and that a moderate or large decline in lung function was associated with a more pronounced decline in HRQoL [14]. Nevertheless, defining relationships between pulmonary function tests and PRO measures remains a significant challenge in ILD.

Hospitalisation is recognised as a serious event by both patients and physicians, and has clear health-economic implications. The frequency of hospitalisations and emergency room (ER) visits was considered important to a CTD-ILD patient focus group [10]. Furthermore, longitudinal data from the German INSIGHTS-IPF registry suggested an association between hospitalisation and poorer HRQoL in patients with IPF [15]. Existing data have demonstrated that patients with SSc-ILD have a higher economic burden compared with patients with other organ involvement associated with SSc, which has been largely attributed to inpatient hospitalisations [3]. In addition, several US-based studies have agreed that SSc-ILD is associated with greater healthcare costs compared with SSc alone [16,17,18]. For patients with incident SSc-ILD, healthcare costs in the USA are nearly twice as high over 5 years compared with patients with incident SSc [16].

In cases where clinically relevant endpoints are difficult to measure or particularly rare in clinical trials, surrogate endpoints such as biomarkers may be a suitable substitute if the two endpoints are known to be correlated. This is also an option for hospitalisation and mortality outcomes in slow, progressive diseases such as SSc-ILD, where large sample sizes and long follow-up times would be required due to low event incidence [19]. A biomarker of SSc-ILD progression such as FVC decline may be able to serve as a surrogate for hospitalisation endpoints if a relationship could be established.

Joint modelling allows for the combined analysis of longitudinal and time-to-event endpoints in a single model and the assessment of the validity of the potential surrogate endpoint [20, 21]. If an association is identified and a treatment effect on the surrogate endpoint is also established, this provides evidence of an indirect treatment effect on the clinically important endpoint [22].

The objective of this post hoc analysis was to investigate the association between longitudinal FVC decline and time to hospitalisation endpoints in SSc-ILD using data from the SENSCIS® trial, and thus to gain insights into the validity of FVC as a surrogate marker for hospitalisation in SSc-ILD.

Methods

Data source

The design and eligibility criteria for SENSCIS® have been reported previously [13]. Briefly, adult patients with SSc [23] were enrolled if they had disease onset (first non-Raynaud symptom of SSc) within the previous 7 years, evidence of lung fibrosis in ≥ 10% of the lung using high-resolution computed tomography, an FVC ≥ 40% of the predicted value, and haemoglobin-corrected DLCO 30–89% of the predicted value. Evidence of previous ILD progression was not required for entry into SENSCIS®. Eligible patients were randomised in a double-blind fashion to receive either nintedanib (n = 288) or placebo (n = 288) for up to 100 weeks or until the last randomised patient reached 52 weeks of treatment, with the primary efficacy evaluation (annual rate of decline in FVC [mL/year]) conducted at 52 weeks. Concomitant treatment with prednisone ≤ 10 mg/day (or equivalent) and/or stable therapy with mycophenolate or methotrexate for ≥ 6 months prior to randomisation was allowed. Additional therapy was allowed during the treatment period if clinically significant deterioration occurred. Patients were stratified according to the presence of anti-topoisomerase I antibody (ATA), which is associated with declining FVC in patients with early SSc [13, 24].

Included time-to-event and longitudinal endpoints

The two outcome variables considered in this study and analysed together in a joint model were defined as follows. Various time to hospitalisation endpoints in the context of the SENSCIS® trial period (over 52 weeks and the whole trial) were considered for inclusion in a joint model. These were (1) time to first all-cause hospitalisation or death, (2) time to first SSc-related hospitalisation or death, (3) time to first admission to the ER or admission to hospital followed by admission to an intensive care unit (ICU) or death, (4) time to first admission to hospital followed by admission to ICU or death, and (5) time to first admission to hospital followed by use of mechanical ventilation or death. Hospitalisation-associated events, including those associated with SSc, were reported and verified by clinicians. Only hospitalisations that occurred during the treatment period or 28 days thereafter were considered.

The longitudinal response variable was the rate of decline in FVC% predicted over 52 weeks and the whole trial. Of note, the rate used here is a model-based estimate that does not require a full observation time of 52 weeks. Only values obtained from on-treatment assessments were considered for inclusion. In a secondary analysis, we further considered the FVC% predicted at a given time (i.e. irrespective of the slope) as the longitudinal response. FVC% predicted was preferred over FVC measured in mL because it already provides an adjustment for prognostic variables, which simplifies the longitudinal modelling.

Statistical analysis

The total sample and subsamples of patients with and without hospitalisation events were characterised using standard descriptive statistics. Joint modelling for longitudinal and time-to-event data was used to assess the association between rate of decline in FVC% predicted and hospitalisation-related endpoints [19, 20]. The approach was implemented using the SAS® macro %JM [25], and validated using the R package JM, version 1.4-8 [26].

Longitudinal sub-model

We used a normal mixed effects model of FVC% predicted with ATA status and FVC% predicted at baseline as predictor variables. Separate mean slopes for patients on nintedanib treatment and controls were assumed, and trajectories were modelled by a linear trend with an unstructured variance–covariance matrix assumed, given the known effect of nintedanib on FVC. All FVC values in the SENSCIS® trial were used: visits at 2, 4, 6, 12, 24, 36 and 52 weeks (visits 3 to 9) for the 52-week time period and visits at 68, 84 and 100 weeks (visits 10 to 12) for whole trial period. Only on-treatment FVC measurements before a hospitalisation event were considered.

Time to event sub-model

For the time to event sub-model, a piecewise exponential model with five internal knots was used to model the baseline hazard. This is a standard approach used in joint modelling of longitudinal and time-to-event data where the baseline hazard is usually modelled explicitly. It assumes that hazards are constant over the particular time intervals and avoids misspecification of time-to-event distribution. FVC% predicted (i.e. the longitudinal response) was used as the endogenous time-dependent covariate, and the time to event sub-model was additionally stratified by ATA status. Of interest was the time to the first event in any of the components of the composite. Patients who experienced an event were censored afterwards and not at risk of further events. This excludes the possibility that events have an impact on the underlying FVC trajectories and thereby potential reverse causality bias.

Association structure

The shared parameter in each of the joint models was FVC% predicted, which was considered as either the “estimated slope” (assuming the rate of FVC% predicted affected the risk of an event occurring; primary analysis) or the “estimated current value” (assuming the risk of a hospitalisation event at a given time depended on the estimated value of FVC% predicted at that time; secondary analysis). Importantly, the estimated slope represents the annual rate of decline in FVC% predicted. It is derived based on the whole estimated trajectory of a patient and allowed to vary over time, and thus is always an estimate referring to a specific point in time. In compliance with recommended reporting, all model coefficients and their precisions are available in Supplementary Methods S1–6 [27].

For each hospitalisation endpoint observed during the SENSCIS® trial with sufficient numbers of events across two time periods (52 weeks and whole trial), joint models were fitted. We present the risk of first hospitalisation events in terms of 1-, 3- and 5-unit FVC decline, as well as graphically from > 0 to 10-unit FVC decline. A “unit” here represents 1 percentage point. Results with a P value ≤ 0.05 were considered statistically significant.

Results

Participant characteristics

A total of 576 participants were enrolled in the SENSCIS® trial and received at least one dose of nintedanib or placebo. FVC and hospitalisation data were available for 574 patients and were therefore included in this analysis. The baseline characteristics of all patients included in the joint models and those with a hospitalisation event or death during the trial period are described in Table 1. Most patients were female (75.1%). Approximately 50% of patients had diffuse cutaneous SSc and 60.8% of the total population were ATA positive. The mean FVC% predicted (± standard deviation [SD]) at baseline was 72.5 ± 16.7 for the total population, 72.0 ± 16.4 and 70.5 ± 18.7 for patients with all-cause and SSc-related hospitalisation events or death and 70.9 ± 16.7 for patients with ER or hospital admission followed by admission to ICU or death. The annual rate of decline in FVC% predicted for patients enrolled in SENSCIS® was − 1.4 ± 0.4% and − 2.6 ± 0.4% for nintedanib and placebo, respectively. Baseline comorbidities and SSc-related organ manifestations, as well as discontinuation data at 52 weeks, are available in Supplementary Results Tables S7–9.

Table 1 Summary of baseline demographics of patients in the SENSCIS® trial
Full size table

Joint model analysis

Six joint models using the estimated slope of FVC decline as the shared parameter were fitted. Low numbers of reported events precluded stable model fits for the endpoints time to first admission to hospital followed by admission to ICU or death (n = 11) and time to first admission to hospital followed by use of mechanical ventilation or death (n = 7).

At 52 weeks, 78 (13.7%) patients either had an all-cause hospitalisation event or died. Of these, 42 experienced a hospitalisation event or death related to SSc, accounting for 7.4% of patients included in the respective joint model. Seventy-five patients were admitted to ER or hospital (including all-cause hospitalisation patients) followed by ICU or death, accounting for 13.1% of patients included in the respective joint model (Table 2).

Table 2 Association between slope of FVC% predicted and risk of first hospitalisation endpoints over 52 weeks
Full size table

We observed a statistically significant relationship between FVC decline and time to first all-cause hospitalisation or death during 52 weeks of the SENSCIS® trial, with a 3-unit decline in FVC% predicted corresponding to a 1.43-fold increase in the risk of an event (hazard ratio [HR] 1.43; 95% confidence interval [CI] 1.24, 1.65; P < 0.0001; Table 2). These values were consistent with those calculated over the whole trial (3-unit FVC% predicted decline: 1.47-fold increased risk; 95% CI 1.25, 1.74; P < 0.0001; Table 3). FVC decline was also significantly associated with time to first SSc-related hospitalisation or death during 52 weeks of the SENSCIS® trial, with a 3-unit decline in FVC% predicted corresponding to a 1.48-fold increase in the risk of an event (HR 1.48; 95% CI 1.23, 1.77; P < 0.0001). Corresponding data for 1- and 5-unit FVC% predicted declines are shown in Table 2. These findings were consistent with those calculated over the whole trial (3-unit FVC% predicted decline: 1.60-fold increased risk; 95% CI 1.29, 1.98; P < 0.0001; Table 3). There was no association found between FVC decline and the time to first admission to ER or admission to hospital followed by admission to ICU or death at either 52 weeks or whole trial (P = 0.1549 and P = 0.3376, respectively). HRs for varying values of FVC decline are visualised in Fig. 1. The estimated slope difference for nintedanib versus placebo in the longitudinal FVC% predicted sub-model was statistically significant at both time points, indicating that the rate of FVC decline was slower with nintedanib (Tables 2 and 3).

Table 3 Association between slope of FVC% predicted and risk of first hospitalisation endpoints over the whole trial
Full size table
Fig. 1
figure 1

Change in risk of first hospitalisation endpoints by decline in FVC% predicted a all-cause hospitalisation or death, b SSc-related hospitalisation or death and c ER or hospital admission followed by ICU or death. Data collected during the treatment period over 52 weeks. CI, confidence interval; ER, emergency room; FVC, forced vital capacity; HR, hazard ratio; ICU, intensive care unit; SSc, systemic sclerosis

Full size image

In additional joint models using the estimated current value of FVC% predicted as the shared parameter (which ignore the individual dynamics), FVC was not statistically significantly associated with the risk of any hospitalisation events or death over 52 weeks (all-cause: P = 0.29; SSc-related: P = 0.13; admission to ER or hospital followed by ICU: P = 0.22; Supplementary Results S10).

Discussion

To the best of our knowledge, this is the first study to assess the impact of FVC changes on the risk of hospitalisation in SSc-ILD. Our joint model approach using data from the SENSCIS® trial showed that a decline in FVC% predicted corresponded to a statistically significant increase in the risk of all-cause and SSc-related hospitalisation or death during the treatment period. This is a new finding in SSc and is consistent with data in IPF [28]; it is also clinically important since hospitalisation rate is one of the most important outcomes for patients. We also observed a statistically significant treatment effect of nintedanib on FVC decline, mirroring the results from the SENSCIS® trial in which nintedanib reduced the rate of FVC decline compared with placebo [13]. Our findings suggest that even small reductions in the decline of FVC% predicted (1–3 units) in response to a treatment intervention can result in relevant reductions in the risk of hospitalisation.

There was no association between the rate of FVC decline and the risk of admission to ER or hospital followed by admission to ICU or death. Although ILD is part of a series of systemic complications in SSc that can lead to hospitalisations, the proportion of ILD-related hospital or ER admissions was not recorded in SENSCIS®, limiting further analysis of this finding. In addition, patients may have required hospitalisation for procedures and special medication administration, but this was also not recorded.

The baseline characteristics of all patients with hospitalisation events were similar to the overall population. Baseline FVC% predicted and DLCO values were numerically slightly lower in patients with hospitalisation events compared with the overall population. Over 20% of hospitalised patients had infections and infestations at baseline, which may be related to concomitant immunosuppression or more severe disease. Furthermore, the trial discontinuation rate was higher in patients who had an all-cause or SSc-related hospitalisation or death than in the overall population. Patients with all-cause or SSc-related hospitalisation events had numerically higher C-reactive protein levels and a higher proportion had elevated creatine kinase levels compared with the overall population. It is possible that patients who had an SSc-related hospitalisation or death may have had more complications related to more active inflammatory disease.

Our analysis used two different parameterisations of the association structure in the joint model. Importantly, we observed strong evidence in the primary analysis that the slope of the FVC decline is predictive of hospitalisation, whereas a secondary analysis exhibited no (or very little) evidence of the current FVC value as a predictor of hospitalisation. As an implication for clinical practice, this suggests that the dynamics of the decline in FVC play a larger role than singular FVC values in the assessment of hospitalisation risk. It also underlines the importance of monitoring changes in FVC over time in clinical practice to assess the hospitalisation risk of individual patients.

Given the potential relationship between the change in FVC and the risk of hospitalisation, we evaluated the validity of FVC as a surrogate endpoint for the clinically important hospitalisation-associated endpoints according to the three levels of surrogacy defined by Taylor and Elston [22]. Since we have observed a statistically significant association between the surrogate outcome of FVC decline and time to hospitalisation events (which is a patient-related outcome), our findings meet the second, intermediate level of the endpoint surrogacy criteria. With an established link between these outcomes, further investigations are required to evaluate whether the slowing effect of nintedanib on FVC decline directly corresponds to a reduction in the risk of a hospitalisation event.

Our study has the following limitations. First, the analysis was limited to the data collected in the SENSCIS® trial, where data on the reasons for hospital admissions were not specific and were limited to whether they were SSc-related or not. A basic linkage with adverse events leading to hospitalisation recorded in temporal proximity to hospitalisations suggests that the following system organ classes were most frequently associated with hospitalisations: “respiratory, thoracic and mediastinal disorders” and “infections and infestations” (each approximately 20% of admissions) and “gastrointestinal disorders” (approximately 10% of admissions). However, these observations need to be treated with caution due to imperfect linkage and missing data. A further limitation is that data collection was restricted to the duration of the clinical trial. Second, the SSc-ILD population had relatively preserved lung function and many were receiving immunosuppressant therapy, which may have slowed down ILD progression but also may have increased the risk for infection-related hospitalisations. Third, as the study was performed in an SSc-ILD population, generalisability to an overall SSc population may be limited. Finally, since these results were obtained from a controlled clinical trial environment, replication in clinical practice would be desirable as the association between FVC decline reduction and risk of hospitalisation reduction is of great clinical importance.

Conclusions

In conclusion, our joint modelling approach indicated that a decline in FVC has a clinically relevant association with the risk of all-cause and SSc-related hospitalisation or death in patients with SSc-ILD in the context of the SENSCIS® trial. It also provides evidence that FVC decline may be used as a surrogate endpoint for time to first hospitalisation, a clinically relevant parameter, though further validation is required. Our analysis further suggests that slowing FVC decline may prevent hospitalisations in patients with SSc-ILD and supports the value of serial FVC measurement in randomised controlled trials.