Evaluation of computer-based computer tomography stratification against outcome models in connective tissue disease-related interstitial lung disease: a patient outcome study

Computed tomography (CT) evaluation of patients with individual connective tissue disease-related interstitial lung diseases (CTD-ILDs) have shown that several parenchymal patterns, including honeycombing [1, 2], reticulation [3], and fibrosis extent, are associated with a poor outcome [1, 4‐6]. However, while studies of prognostic indices within individual CTDs convey valuable information about specific, small patient groups, the applicability of such indices to a wider group of “all-comers” CTDs needs validation.

The importance of identifying prognostic indices across a population of various CTD diagnoses lies in the fact that CTD sub-groups often overlap both in their clinical and CT characteristics. Yet, there are very few CT studies that have considered mixed populations of CTD patients. One such study, by Walsh et al. [7], identified severity of traction bronchiectasis and honeycombing as indices predictive of mortality, confirming the importance of two parenchymal patterns previously shown to be prognostically important in the non-CTD idiopathic interstitial pneumonias [8‐10].

Computer-based CT analysis in the CTDs [11, 12] has been relatively neglected when compared to idiopathic pulmonary fibrosis (IPF) [13‐15]. Furthermore, the application of advanced mathematical modelling techniques to CT datasets has been limited thus far [16] despite the modelling of quantified CT variables having the potential to provide a comprehensive morphological analysis of a patient’s disease. By evaluating the entirety of a CT dataset, computer tools, when allied to modelling techniques, can identify patient clusters that share similar disease phenotypes and potentially identify sub-groups with similar outcomes.

The current study therefore compared the strength of visual and computer-based (CALIPER) CT patterns and pulmonary function tests (PFTs) for the prediction of mortality for a mixed cohort of CTD-ILD patients. A secondary analysis evaluated mortality prediction across the entire cohort using mathematical modelling of CALIPER-scored CT variables and compared mortality prediction against the interstitial lung disease gender, age physiology (ILD-GAP) outcome model.

Our study has demonstrated for the first time, that, across the range of CTD-ILD diagnoses, a computer-derived CT parameter, the pulmonary vessel volume, is an independent predictor of mortality. Furthermore, the PVV is a stronger predictor of mortality than all other CT and pulmonary function variables following correction for age and gender. In addition, automated stratification of CALIPER-derived CT variables identifies patient groups with distinct characteristics, and three automated stratified groups demonstrated significantly different functional profiles and patient outcomes. When the functional indices (FVC, DLco) in the ILD-GAP model were substituted with the automated stratified groups, the new Stratified CT model improved mortality prediction when compared to the ILD-GAP model. When the automated stratified groups were subsequently combined with the ILD-GAP model (Stratified CT-GAP model), mortality prediction was further augmented. Accordingly, automated stratified CALIPER CT variables have the potential to be used as an alternative to, or combined with, functional indices to predict outcome in CTD-ILD patients.

Our observations are particularly relevant given a recent editorial which articulated the need to improve the identification of distinct disease phenotypes in patients with rheumatoid arthritis related-ILD to aid risk prediction and diagnosis [30]. Apart from systemic sclerosis, most studies in the CTD-ILDs have been constrained by small patient numbers. Accordingly, there is a growing need to combine patient cohorts across centres to generate more substantial and inclusive datasets [30]. Although CT evaluation is near ubiquitous in the setting of known or suspected CTD-ILD, the complexities and inconsistencies associated with visual CT scoring demand more robust alternatives for the quantification of disease patterns and extents.

Computer analysis of CTs in CTD-ILD populations is an attractive alternative to visual scoring and when combined with the unbiased nature of automated stratification, may allow the identification of patient phenotypes that are visually subliminal. In addition, the strength of DLco as a predictor of outcome in CTD-ILD may well be diminished in multicentre cohorts given the variation associated with DLco measurements across laboratories [31], further emphasising CT evaluation as a potential outcome measure in patients with CTD-ILD. In this regard, the ability of automated stratification to substitute for DLco and FVC measured at a single institution, without loss of strength in outcome prediction, argues for consideration of computer-based CT analysis in future multicentre CTD-ILD studies.

The improved strength of the Stratified CT-GAP model over the ILD-GAP model identified in the current study is largely a consequence of a confounding effect of the normal range of pulmonary function when PFTs are stratified as thresholds. The range of normal pulmonary function values extends across the range of 80–120% of predicted values based on patient age, gender, race and height. As a result, in a staging system, if a patient lies close to a lung function threshold, small differences in predicted normal values will have a major impact on how the patient is staged, shifting them above and below thresholds. For example, if a patient started with a predicted FVC at 120% and lost 35% of predicted lung function they would remain as GAP stage 1. However, if the patient started at a predicted FVC of 80% and lost 35% of predicted FVC, they would fall into GAP stage 3. Consequently, the normal range has a dramatic effect on the severe end of the spectrum of disease in determining where someone lies on the GAP scale.

A similar limitation of a “normal range” is not present in morphological CT variables however and CT variables can therefore serve to modify confounding effects associated with clustering around PFT thresholds as identified in a previous study evaluating a scleroderma staging system [6]. Goh et al. [6] showed that threshold measures of CT (Hazard ratio [HR] = 2.5) and PFT (HR = 2.1 for an FVC threshold) variables were significantly weaker when analysed alone, but improved considerably when structure and function were combined (HR = 3.5). Similarly, in the current study, the ILD-GAP model was a less sensitive predictor of mortality secondary to the clustering of individuals around PFT thresholds, an effect that was partially ameliorated following amalgamation of the Stratified CT score to the ILD-GAP model.

The current study is the first of its kind to evaluate mortality prediction in CTD patients using computer-based volumetric CT analysis. Several previous studies in CTD patients analysing CT scans with computer algorithms have utilized interspaced high-resolution CT imaging [11, 12], precluding the robust evaluation and differentiation of patterns such as honeycombing and emphysema. The remaining computer-based studies have evaluated the lung according to its simple density characteristics, deriving metrics of histogram skewness and kurtosis [32‐34]. Such metrics have been shown to correlate poorly with other markers of disease severity and with mortality in IPF [15] and are relatively unsophisticated compared to modern structural and textural analytic techniques [27]. Furthermore, only the studies by Marten et al. [32, 33] evaluated computer scores against physiological indices whilst the remaining studies compared computer-based scores with visual CT scoring. No studies to date have evaluated computer scores against mortality in patients with CTD.

CALIPER has advantages over most quantitative tools by virtue of its volumetric structural and textural analysis of the lung, which, for example, enables low attenuation areas of the lung to be distinguished as representing either honeycombing or emphysema [27]. Similarly, volumetric analysis allows quantitation of features that cannot be resolved visually, such as the percentage of the lung volume composed of vessels [25]. A glyph distils CALIPERs quantitative data into a format that is easily deciphered by the non-specialist in a busy clinic setting, which may have crossover utility for both rheumatologists and pulmonologists in the evaluation of patients with CTD-ILD. Whilst the glyph presentations are a by-product of CALIPER analysis we do not wish to give undue prominence to them in the current study however, since it is based on population characteristics rather than individual patient/glyph appearances. Interrogating an individual glyph, which simplifies complex spatial patterns of disease morphology and extent, is an inferior exercise when compared to the modelling analyses conducted in the current study. To derive absolute conclusions about an individual’s likely outcome based solely on a glyph would be misleading.

There are very few large-scale studies that have evaluated the ability of CT variables to predict mortality across all CTD subtypes. A study by Walsh et al. [7] evaluated CTs and pulmonary function indices in 168 patients with various CTDs and found that traction bronchiectasis severity and honeycombing extent scored visually along with DLco were independently predictive of mortality. In the present study, across all CTD-ILD patients, when visually scored CT parameters were analysed alone, visual honeycombing and traction bronchiectasis severity scores were also independently predictive of mortality. However, when combined with CALIPER CT variables and PFTs, however, visual honeycombing and traction bronchiectasis scores did not retain prognostic significance.

The association between pulmonary hypertension and connective tissue diseases has long been recognised [35], and supervening pulmonary hypertension is associated with a poor outcome across the range of CTDs [36‐38]. It would therefore be logical to assume that the mortality signal associated with PVV reflected a new imaging marker of pulmonary hypertension. However, as with our observations in patients with idiopathic pulmonary fibrosis [25, 39], we identified only weak linkages between PVV and both RVSP and Kco. Indeed, Kco and PVV were independently predictive of mortality across the range of CTD patients. The findings suggest that the PVV signal does not primarily reflect the severity of pulmonary hypertension, or indeed act as a key marker of damage to the vascular compartment of the lung. The counter-intuitive relationship between PVV and the extent of ILD identified on CT may, as previously postulated, be explained by local increased vascular pressures within fibrotic regions of the lung that result in blood diversion to spared lung regions. As fibrosis worsens and vessel size and number (above a size threshold recognised by CALIPER) increase in non-fibrotic regions of the lung, the accompanying increase in CALIPER PVV may effectively act as a surrogate marker of ILD extent [39].

The superiority of PVV in predicting mortality over CALIPER and visually scored total ILD extents may relate to the specific pathophysiologic changes that develop in the lung secondary to fibrosis. As fibrosis worsens, the lung contracts with the result that the extent of fibrosis, when measured volumetrically or expressed as a proportion of the total lung volume may, in fact, decrease. Consequently, in a patient with more severe disease, a volumetric CT score of fibrosis extent underestimates fibrosis severity. PVV avoids such a pitfall, as it is a parameter that increases in line with fibrosis extent. Evaluation of PVV as a prognostic marker in the fibrosing lung diseases remains in its infancy; however, results from the current study argue for further detailed study of the variable in other fibrosing lung diseases as well as evaluation of PVV as a marker of deterioration on serial CT evaluation.

There are some limitations to this retrospective study. Firstly, the individual CTD-ILD diagnoses making up the cohort were not evenly distributed, for example, there were large numbers of RA-ILD and systemic sclerosis-ILD patients but very few cases of SLE. In mitigation, however, given that the study population represented a consecutive cohort of new clinic presentations, the case mix arguably represents a real-world caseload. Secondly, there were only 14 patients in automated stratified group 1, limiting the strength of statistical relationships between groups. A consequence was that the remaining patients were split into two groups generating dichotomous good and bad outcome groups. Since most management decisions are binary with regard to giving or withholding medication, a two-group model is usually preferable to a multiple group model where managing patients in intermediate outcome groups is problematic. It could also be argued that some patients with an apparently good outcome may turn out to have a delayed poor outcome once treatment benefits have dissipated. Such a reservation is common to a great many studies, and yet primarily evaluating all patients at presentation does, at least, provide a satisfactory spread of disease severity, including some patients with earlier disease and others with more advanced disease. There are limitations associated with making exact prognostic separations based on baseline evaluation and conclusions reached at a single point in time should retain some flexibility to enable modification by observed changes in subsequent disease behaviour. Finally, an external validation cohort would ideally have been used to confirm our findings; however, the scarcity of large, well characterised fibrosing lung disease cohorts remains a recognised constraint.

In conclusion, we have demonstrated that, in a large mixed population of CTD-ILD patients, CALIPER pulmonary vessel volume was the CT variable that best predicted mortality and may be a new prognostic index. When automated stratified CALIPER variables were substituted for the functional indices in the ILD-GAP index, mortality prediction was strengthened. Computer analysis and automated stratification of CTs may therefore represent a viable alternative to visual CT scoring and evaluation of functional indices in patients with CTD-ILD, and demonstrates added value when combined with outcome prediction models such as the ILD-GAP model.