e-Lung Computed Tomography Biomarker Stratifies Patients at Risk of Idiopathic Pulmonary Fibrosis Progression in a 52-Week Clinical Trial

To the Editor:

Since the successful pivotal studies of pirfenidone (1) and nintedanib (2) for idiopathic pulmonary fibrosis (IPF), several drugs have shown promise in early-phase clinical trials (3, 4) only to fail on efficacy grounds in registrational studies (5). With antifibrotic therapy representing standard of care, successful trials rely on demonstrating further attenuation in FVC decline. Enrolling patients who remain stable over the course of a clinical trial represents a risk, as progression is necessary to be able demonstrate a treatment effect. Furthermore, slowly progressive control arms potentially represent an insurmountable challenge for a novel therapeutic. There is, therefore, a need to identify novel means by which future clinical trials can be enriched for patients who are more likely to progress and to ensure that treatment arms are well matched for likelihood of progression.

e-Lung (Brainomix) is a machine learning–developed, automated computed tomography (CT) processing algorithm, that quantifies features relevant to interstitial lung diseases (ILDs). The weighted reticulovascular score (WRVS) (6, 7) is an e-Lung quantitative CT biomarker that incorporates reticular opacities and vascular structures in the periphery of the lung, expressed as a percentage of the lung rind.

In the Phase 2 clinical trial of tralokinumab (ClinicalTrials.gov ID: NCT01629667), patients with IPF received 400 mg, 800 mg, or placebo every 4 weeks for 68 weeks. The study was discontinued early on efficacy grounds, with a mean absolute FVC decline of −160 ml in the placebo arm, −210 ml in the 400-mg dose, and −210 ml in the 800-mg dose at 52 weeks (8). Patients who enrolled in the study had baseline but no subsequent CT scans.

In this post hoc analysis, we evaluated the association between baseline WRVS and time to FVC decline of ⩾10% from baseline by 52 weeks. Regardless of treatment allocation, in this analysis, patients’ FVC declined by broadly the same magnitude (for placebo, 136 ml; for 400 mg tralokinumab, 170 ml; for 800 mg tralokinumab, 186 ml; ANOVA P = 0.43); analyses were, therefore, handled as a single cohort. Patients were included if they had a routine baseline CT scan within 12 weeks of first dosing. All patients had FVC measurements at enrollment, at intervals throughout the study, and at 52 weeks. Patients were divided into low- and high-risk groups using clinically relevant (FVC, ⩾80% and <80%, respectively; Dl_CO, ⩾40% and <40%, respectively; and a GAP [gender, age, physiology] Index of 1 and 2–3, respectively) and clinically applicable WRVS thresholds that represented the median (12.3%)—an approach utilized in previously published work (9)—rounded to the nearest 5% thresholds (10% and 15% for low and high risk, respectively). Cox regression (hazard ratios [HRs] with 95% confidence intervals [CIs]) was used to analyze time to FVC decline. Univariable and multivariable analyses are presented. Harrell’s C-indices are reported to compare the strength of association between baseline variables and outcomes, where 0 indicates a perfect negative association; 0.5, being random; and 1, a perfect positive association.

Sixty-two patients were analyzed; the mean age was 67 years, and 50 (81%) were male. In a univariable analysis using WRVS (%), FVC (%), and Dl_CO (%) as continuous variables, WRVS had the strongest association with FVC decline at 52 weeks (C-index of 0.77, compared with FVC [0.61] and Dl_CO [0.72]). A WRVS threshold of ⩾15% identified patients at high risk of FVC decline with a HR of 5.74, (95% CI: 2.07, 15.92), P < 0.05 (Figure 1), compared with a baseline Dl_CO < 40% (HR, 3.54; 95% CI: 1.28, 9.79; P < 0.05), GAP Stage 2–3 (HR, 2.59; 95% CI: 0.88, 7.58; P = 0.08), and baseline FVC < 80% (HR, 1.29; 95% CI: 0.41, 4.06; P = 0.66). Patients with a WRVS of <10% were at low risk of 10% FVC decline with a HR of 0.2 (95% CI: 0.03, 1.52), P = 0.12 (Figure 2). We evaluated whether combinations of continuous variables could improve associations. A regression analysis with WRVS and age was strongly associated with 52-week FVC decline, with a C-index of 0.76; this was further improved by adding baseline FVC percentage to a C-index of 0.81. Finally, we evaluated the relationship between baseline WRVS, FVC, and Dl_CO with annualized FVC percentage change. Baseline WRVS had the strongest relationship with FVC decline, with an R² value of 0.23, compared with FVC (R² = 0.14) and Dl_CO (R² = 0.12).

Figure 1.

Kaplan-Meier curve showing stratification of high (⩾15%) WRVS and lung function decline as defined by relative drop in FVC of ⩾10%. WRVS = weighted reticulovascular score.

Figure 2.

Kaplan-Meier curve showing stratification of low (<10%) WRVS and lung function decline as defined by relative drop in FVC of ⩾10%. WRVS = weighted reticulovascular score.

We have previously demonstrated that the WRVS is prognostic, independent of lung function and radiologist assessment of fibrosis extent (6). In this post hoc analysis, we have shown the potential in a clinical trial setting for the e-Lung WRVS biomarker to exclude patients at low risk of progression and to enrich for patients at high risk of progression. By Week 52, only 1 patient in the low-WRVS group had experienced a decline in FVC ⩾ 10%, in contrast to 14 in the high-risk group. There is a precedent for excluding patients at low risk of progression in clinical trials (e.g., the INBUILD and ZEPHYRUS1 studies, which excluded patients with <10% visually assessed fibrosis) and for enriching for those at the highest risk of progression, such as the approach that was used in the INBUILD study (10), where there was enrichment for a usual interstitial pneumonia-like phenotype on CT. By demonstrating that baseline WRVS has the strongest relationship with FVC change at 52 weeks, we have demonstrated its utility as a prognostic biomarker. With e-Lung incorporated into a clinical trial design, and with WRVS as a covariate, the sample size for a 52-week study with change in FVC percentage as the primary outcome could be reduced by 23% by lowering the uncertainty around a treatment effect estimate. It is also notable that an impaired Dl_CO was more prognostic than a low FVC, and this could be taken into consideration when designing future IPF clinical trials.

Although this study was limited by sample size, given that only 62 of the 103 patients completing the 52-week study had a CT scan within 12 weeks of first dosing, the strength of the associations observed are robust and in line with our a priori hypothesis, given previously presented work (6, 7). As this study was conducted in the periantifibrotic era, we could not adjust for use of antifibrotic therapy. This will be important in being able to generalize results to future prospective clinical trials. Notwithstanding this, in the recent ISABELA clinical trials (5), the magnitude of FVC decline in patients treated with an antifibrotic mirrored the placebo arms of the pivotal studies of pirfenidone (1) and nintedanib (2). The evolving nuances of the clinical trial landscape demonstrate the value of utilizing novel artificial intelligence–based techniques to adjust for confounders.

In this post hoc analysis of an IPF clinical trial, we have shown that WRVS is associated with FVC decline at 52 weeks, outperforming standard measures. These data suggest that the e-Lung WRVS tool may allow for enrichment of clinical trials with progressive patients, could identify patients at low risk of IPF progression, might facilitate well-matched treatment arms, and could reduce the size of future clinical trials. Further work is needed to validate these findings in additional cohorts.