Clinical and Molecular Findings After Autologous Stem Cell Transplantation or Cyclophosphamide for Scleroderma: Handling Missing Longitudinal Data

Abstract

Objective: Among individuals with systemic sclerosis (SSc) randomized to cyclophosphamide (CYC) (n = 34) or hematopoietic stem cell transplantation (HSCT) (n = 33), we examined longitudinal trends of clinical, pulmonary function, and quality of life measures while accounting for the influence of early failures on treatment comparisons.

Methods: Assuming that data were missing at random, mixed-effects regression models were used to estimate longitudinal trends for clinical measures when comparing treatment groups. Results were compared to observed means and to longitudinal trends estimated from shared parameter models, assuming that data were missing not at random. Longitudinal trends for SSc intrinsic molecular subsets defined by baseline gene expression signatures (normal-like, inflammatory, and fibroproliferative signatures) were also studied.

Results: Available observed means for pulmonary function tests appeared to improve over time in both arms. However, after accounting for participant loss, forced vital capacity in HSCT recipients increased by 0.77 percentage points/year but worsened by -3.70/year for CYC (P = 0.004). Similar results were found for diffusing capacity for carbon monoxide and quality of life indicators. Results for both analytic models were consistent. HSCT recipients in the inflammatory (n = 20) and fibroproliferative (n = 20) subsets had superior long-term trends compared to CYC for pulmonary and quality of life measures. HSCT was also superior for modified Rodnan skin thickness scores in the fibroproliferative subset. For the normal-like subset (n = 22), superiority of HSCT was less apparent.

Conclusion: Longitudinal trends estimated from 2 statistical models affirm the efficacy of HSCT over CYC in severe SSc. Failure to account for early loss of participants may distort estimated clinical trends over the long term.

Figure 1:. Forced Vital Capacity (FVC) trajectories for individuals with reasons for termination

Each line connects FVC values (% predicted) over time for an individual participant. Line colors and symbols at the last assessment indicate reasons for termination: completed study per protocol (blue triangle), death (black star), early withdrawal (red triangle), early last FVC (blue circle), organ failure (gold diamond). The dashed black line connects means (black asterisks) for available data clustered within the following time intervals in months: midpoint (range): 0, 8 (5–11), 14 [11–17), 20 [17–23), 26 [23–29), 32 [29–35), 38 [35–41), 44 [41–47), 48 (47–51), 54 [51–57), 60 [57–63), 66 [63 and above]. Numbers of subjects available for each interval are given above the x-axis.

Figure 2:. Estimated longitudinal trends for mixed effects regression and shared parameter models

Estimates for the hematopoietic cell transplant and cyclophosphamide arms are presented in blue and red, respectively. Solid lines show model-based estimates for mixed effects regression models; vertical lines show 95% confidence intervals at select time points. Dashed lines show model-based means for shared parameter models. Dotted lines connect means for available data clustered as per Figure 1. Figure 2A, Forced Vital Capacity (FVC); 2B, Diffusing Capacity of the Lung for Carbon Monoxide (DLCO); 2C, modified Rodnan Skin Score (mRSS); 2D, Health Assessment Questionnaire-Disability Index (HAQ-DI).

Figure 3.. Trends from mixed effects regression models for intrinsic molecular subsets: Forced Vital Capacity (FVC)

Estimates for the transplant and cyclophosphamide arms are presented in blue and red, respectively. Solid lines show model-based estimates for mixed effects regression models; vertical lines show 95% confidence intervals at select time points. Mixed effects regression models are as described for Figures 2A with the addition of separate intercepts and slopes (or decay rates) for each treatment-by-intrinsic-subset group.

Figure 4.. Trends from mixed effects regression models for intrinsic molecular subsets: Diffusing Capacity of the Lung for Carbon Monoxide (DLCO)

Estimates for the transplant and cyclophosphamide arms are presented in blue and red, respectively. Solid lines show model-based estimates for mixed effects regression models; vertical lines show 95% confidence intervals at select time points. Mixed effects regression models are as described for Figures 2B with the addition of separate intercepts and slopes (or decay rates) for each treatment-by-intrinsic-subset group.