Eight19 (United Kingdom)

<bold>Introduction:</bold> Pulmonary arterial hypertension (PAH) patients are increasingly older with comorbidities, leading to management and prognostic implications. Deep learning-based segmentation of CTPA with quantification of pulmonary blood volumes and parenchymal abnormalities may help these patients. <bold>Methods:</bold> We used a deep learning model to quantify pulmonary blood and ground glass (GGO) volumes normalised to lung volumes on CTPA in a secondary analysis of the prospective 2000-2024 CAPHTURE (Cambridge PH Registry) PAH cohort. Volumes Z-scores were separated into tertiles, data reported in medians and survival analysed with Kaplan-Meier and Cox regression. <bold>Results:</bold> 120 PAH patients (median age 50, 80% female) were recruited. Patients in the higher GGO/lung volume tertile had higher number of cardiac comorbidities (0 vs 1 vs 1, p=0.01), reduced 6-minute walk distance improvement after PAH medication (54 vs 12 vs 12m, p=0.002), poorer tolerability of optimal PAH medications (p<0.001), required more diuretics (p<0.001), and had poorer survival (Figure 1). Patients in the highest vein/lung volume tertile had more cardiac comorbidities (0 vs 1 vs 1, p=0.048) and poorer survival. <bold>Conclusion:</bold> Automated quantification of GGO and pulmonary vein volume from CTPA can help management decisions by phenotyping PAH patients with cardiometabolic comorbidities who are less likely to benefit from current PAH medications and have worse survival. <fig><object-id>erj;66/suppl_69/PA6175/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>

<bold>Objectives:</bold> To generate a synthetic control arm (SCA) for efficacy testing of buloxibutid in idiopathic pulmonary fibrosis (IPF). <bold>Methods:</bold> AIR was an open-label, phase 2a trial of oral buloxibutid 100 mg BID for up to 36 weeks in treatment-naïve IPF patients. For SCA, 20,000 control arms (using real world data from the Qureight platform) were generated by Monte Carlo Cross Validation (MCCV). Matching similarity was assessed using chi-squared, t-, and kernel tests. Matching was on age, sex, FVC%, DLCO%, and FEV1%. 408 randomly sampled groups of 48 patients were created. Discontinuations from the treatment arm were imputed according to the mean SCA decline. <bold>Results:</bold> SCAs were well matched with the treatment arm (p=0.23). SCAs generated reference distributions of mean FVC (mL) change (Fig 1A). Mean change in FVC for the treatment group after 36 weeks was 23.2ml (SD=508.2, N=48) vs. -114.8ml (SD=45.2) across 408 SCAs (Fig 1B). The right-tailed test for treatment effect (a=0.05) based on this distribution yielded p-value of 0.0025 or expressed alternatively, 99.75% of SCAs had mean FVC change less than the AIR cohort. <bold>Conclusions:</bold> We have demonstrated a statistically significant treatment effect for 100mg BID of buloxibutid at 36 weeks, utilizing MCCV-generated synthetic control arms with matched clinical variables. This is the first use of synthetically generated control arms for efficacy testing in IPF. <fig><object-id>erj;66/suppl_69/PA3043/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig>

Abstract Rationale: The prognostic separation between patients with ‘UIP-like disease’ and’ other fibrotic patterns’ on CT in IPF and progressive pulmonary fibrosis is well established. Several studies also emphasise the importance of short-term disease behaviour over radiologic diagnosis alone. We investigated the predictive utility of deep learning-based CT phenotypes (SOFIA) and longitudinal change in quantitative CT metrics (Lung8, Air8, Vascul8, and Fibr8) in patients with IPF. ---Methods: Participants from PROFILE (a multicentre, prospective observational cohort of IPF patients) were evaluated. SOFIA provided stratification of patients using the INBUILD stratification by baseline CT phenotype (UIP-like disease [UIP and probable UIP] and other fibrotic patterns [indeterminate or alternative diagnosis). Airway volume, vascular volume and fibrosis volume (expressed as a % of lung volume) were quantified using Air8, Vascul8, and Fibr8, respectively. Cox regression and Harrell's C-index were used to investigate associations between algorithm outputs and mortality. Results: 417 Patients had suitable baseline CTs. SOFIA-based stratification was UIP-like (n=334) and other fibrotic patterns (n=83). UIP-like disease was associated with increased mortality, independent of baseline Fibr8 extent (HR 1.46, p=0.013, 95%CI 1.08-1.99, C-index 0.67). Of the CT metrics, baseline fibrosis (Fibr8) and airway (Air8) volumes were both associated with increased mortality in UIP-like disease: HR 1.06 p&amp;lt;0.0001 95%CI 1.04-1.07, HR 2.20, p=0.001 95%CI 1.37-3.55, C-index 0.68 while only Fibr8 extent was associated with increased mortality in patients with other fibrotic patterns: HR 1.04 p&amp;lt;0.0001 95%CI 1.02-1.08, C-index 0.66. 134 patients had at least one follow-up CT (mean follow-up period 570±363 days). On multivariate analysis, baseline Fibr8 extent and annualised Fibr8 change were independently associated with increased mortality while baseline INBUILD categorisation was not (HR1.08 p&amp;lt;0.0001 95%CI 1.05-1.10, HR 1.05 p=0.002 95%CI1.02-1.08, HR 1.31 p=0.288 95%CI 0.80-2.19, C index 0.71). Conclusions: Fibrosis progression on CT is a stronger predictor of mortality in IPF than baseline guideline-based CT diagnosis. For prognostication, emphasising precise measurement of short-term changes in fibrosis on CT may be more effective than classifying baseline CT appearances according to current guidelines.

Abstract Rationale: Most outcome studies in idiopathic pulmonary fibrosis (IPF) using computer-based CT analysis have focused on either phenotype or measures of disease extent in isolation. We evaluated the combination of baseline SOFIA, a UIP classifier and deep learning-based segmentation algorithms, Lung8, Air8, Vascul8, and Fibr8, for outcome prediction in a prospective IPF cohort. Methods: Participants from PROFILE with CT scans suitable for analysis were included. SOFIA provided guideline-based radiologic diagnoses and a probability score for usual interstitial pneumonia (UIP). Lung volume, airway volume, vascular volume and fibrosis volume were quantified using Lung8, Air8, Vascul8, and Fibr8, respectively. Survival analysis was performed using Cox regression. Optimal thresholds were derived for predictive metrics using ROC curve analysis. Two independent sites contributed to PROFILE; Site A was used as a derivation cohort, and Site B as a replication cohort. Results: 417 patients in PROFILE (Site A=247, Site B=170) had baseline CT scans suitable for analysis. Air8, Vascul8, and Fibr8 outputs were normalised using Lung8 (total lung volume). For Site A, SOFIA %UIP probability and normalised-Fibr8 extent (%fibrosis extent) were independently associated with increased risk of mortality (HR 1.01, p&amp;lt;0.0001, HR 1.04 p&amp;lt;0.0001, respectively, C-index=0.70). SOFIA UIP probability and normalised-Fibr8 extent had low correlation (r2=0.28 p&amp;lt;0.0001). A 3-stage risk score was created using thresholds for SOFIA UIP probability (22.2%, AUC 0.71) and normalised-Fibr8 extent (15.3%, AUC 0.74), optimised against disease progression stratifying patients into low (below threshold in both categories), moderate (above the threshold in one category), and high-risk (above threshold in both categories) groups. This risk score was associated with increased mortality at Site A (HR 2.21, p&amp;lt;0.0001, C-index 0.67) and replicated in Site B (HR 2.00, p&amp;lt;0.0001, C-index 0.64) and was the only metric associated with increased mortality on bivariable analysis with each threshold in both cohorts. In the whole cohort, increasing risk score was associated with increased risk of mortality controlling for %predicted FVC (HR 1.84, p&amp;lt;0.0001, C-index 0.69), %predicted DLco (HR 1.38, p=0.001, C-index 0.73) or the composite physiologic index (HR 1.37, p=0.002, C-index 0.72). Conclusions: CT phenotype and fibrosis quantification are independently associated with mortality in IPF and can be combined in a 3-stage risk score for predicting outcomes. Figure 1 KM Survival curve depicting survival distinctions based on the 3-stage risk score in Site A and Site B (n=417)

The PINTA trial (NCT03725852) was a double-blind placebo (PBO)-controlled study of GLPG1205 in 68 IPF patients. At 26 weeks, patients randomised to GLPG1205 showed a numerically smaller decline in FVC (primary outcome). We explored use of&nbsp;automated CT analysis as an alternative endpoint. CT scans were uploaded to the Qureight platform to allow real-time quantification and longitudinal assessment of change in lung (LUNG8) and airway (AIR8) volumes.¹ 43 patients with at least two scans 26 weeks apart were included for analysis. After quality control, 42 were suitable for LUNG8 and 41 for AIR8. Statistical analysis was performed with python, power analysis with R-package Superpower. A significant difference in lung volume change over 26 weeks was found (GLPG1205=+1.51% vs PBO=–4.48%, p=0.027). There was no significant difference in standardised airway volume change between groups (GLPG1205=–0.57% vs PBO=0.26%, p=0.85), Fig 1. The estimated effect size in lung volume (0.71) implies that n=36 per group is sufficient to achieve 80% power. Our data further support that GLPG1205 slows reduction in lung volume over time compared to PBO in IPF. Novel imaging quantification software may provide alternative trial endpoints allowing for smaller trial sizes. These will require standardised imaging protocols and power calculations. 1. Roberts M. ERS 2021, OA3951 <b>Funding:</b> Galapagos, Qureight