Background/Purpose:
Interstitial lung disease (ILD) is a common cause of mortality in systemic
sclerosis (Ssc). Decreased forced vital capacity
(FVC) in Ssc-ILD is associated with increased
dyspnea, is predictive of more rapid development of clinically significant
pulmonary fibrosis, and is associated with increased mortality. Methods are
needed to prospectively identify patients at risk for progressive ILD to guide
clinical trial enrollment and early treatment.

Methods:
Continuous curves indicating the most likely future FVC values of a given
individual are predicted using clinical data (see Fig. 1). Our model uses
individual characteristics including demographic (gender and race) and
serologic measurements (ACA and Scl-70 antibody positivity). In addition, it
uses the FVC history of the individual to dynamically update predictions. Key
to our approach is the idea of disease subtypes: subgroups of individuals with
similar lung disease trajectories [1]. These trajectories are derived with a
statistical learning algorithm (see for example [2]) using prospectively
collected data. Individual-specific model parameters further personalize
predictions. Our algorithm is being adapted to a web-based interface, allowing
practitioners to obtain personalized estimates of a patient’s future disease
course.

Results:
Figure 1 shows two individuals who test positive for Scl-70 antibodies and
diffuse skin disease—both typically associated with development of
progressive ILD. The two most probable subtypes are shown. Baseline FVC levels
are comparable across both patients. After one year of followup,
indicated by points (A) and (B), our model is able to correctly predict that
the FVC in the individual in the first row will remain stable, while the other
will experience FVC decline. After a transient decrease in FVC in the top
patient resulting from an episode of cholecystitis,
the predicted FVC trajectory appropriately continued to predict a largely
stable course of ILD. After 4 years of data the confidence in each prediction
is strengthened in spite of the sharp consecutive drop that both individuals
exhibit­—indicated by points (C) and (D). Using 10-fold cross
validation on 672 patients, our model achieves mean absolute errors of 10.37,
8.95, and 6.98 when predicting pFVC values between
8-12 years of followup using 1, 2, and 4 years of
data respectively.

Conclusion:
Our prognostic model allows dynamic prediction of an individual’s FVC
trajectory from limited initial demographic and serologic data, and could
inform recruitment of patients with similar expected ILD course into clinical
trials.  Incorporation of the model
into a web-based decision-support tool could provide real-time prognostic data
for widespread clinical practice.

References:

   
[1]  Saria
and Goldenberg, IEEE Intelligent Systems, 2015

    [2]
Schulam et al., Proc. of AAAI, 2011