Background:Management of patients with systemic sclerosis-associated interstitial lung disease (SSc-ILD) is complicated by high inter-patient variability. To date, no validated predictors of treatment response are available for routine use. High resolution computed tomography (HRCT)-based radiomics, i.e. the high-dimensional, quantitative analysis of imaging metadata, have previously been shown to be successful in discriminating (SSc-)ILD phenotypes in preclinical and clinical studies1. Since HRCT is an integral part of the routine work-up in SSc, HRCT-based radiomic features may hold potential as non-invasive biomarkers.Objectives:To predict treatment response using two-dimensional (2D) HRCT-based radiomics in SSc-ILD patients from a prospectively followed cohort.Methods:Inclusion criteria were diagnosis of SSc-ILD in HRCT, availability of a suitable chest HRCT scan within 12 months prior to initiation of a new treatment, and availability of clinical baseline and follow-up information. Treatment response was defined as the absence of all of the following over a follow-up period of 12-24 months: relative decrease in forced vital capacity (FVC) ≥5%, increase of ILD in HRCT as assessed by a radiologist, change in treatment regimen due to insufficient response, ILD-related death or lung transplantation. Of each pre-treatment HRCT, 6 slices (15±5 mm apart, starting from the basal lung margin) were manually segmented and 1513 2D radiomic features were extracted using the in-house software Z-Rad (Python 2.7). Features were Z-score transformed and pre-filtered for inter- and intra-reader robustness (intraclass correlation coefficient >0.85) and inter-feature correlation (Spearman’s rho <0.9). A categorical linear regression model was created using 3-fold cross-validated elastic nets for feature selection. Features were then summarized and divided by their number. For generation of a score cut-off, Youden’s score was used. For two-group analyses of continuous variables, Wilcoxon’s test was performed, whereas categorical data was assessed using Fisher’s exact test.Results:A total of 64 pre-treatment HRCTs from 54 patients were analyzed. In 9 patients, >1 asynchronous treatments were assessed, while 45 patients had only 1 eligible treatment approach. The response rate within the assessed follow-up period was 45.3% (n=29). For score generation, 13 radiomic features were selected and an optimal cut-off value of -0.1589 was determined. Univariate linear regression showed significant association between our categorical radiomics-based score and treatment response (p=0.007, area under the curve = 0.65 (0.51-0.79), sensitivity=0.90, specificity=0.43), whereby a high score was predictive for treatment response.No differences between patients with high (n=46) or low (n=18) scores were detected for baseline age (mean±SD=55.5±12.0 and 55.5±13.6 years, p=0.84), duration of SSc (mean±SD=6.2±8.4 and 4.7±4.4 years, p=0.79), time since ILD diagnosis (2.7±2.9 and 2.4±3.1 years, p=0.59), FVC (77.6±20.6 and 80.1±17.9, p=0.41) or DLco (54.4±21.0 and 57.6±18.9, p=0.40). Distribution of anti-Scl-70 positivity (45.7% vs. 55.6%, p=0.58) and diffuse cutaneous disease (47.7% vs. 61.1%, p=0.41) was not significantly different between patients with high and low scores, respectively, although a trend towards higher percentages in the high score group was observed.Conclusion:Our results indicate that, following validation in external cohorts, radiomics may be a promising tool for future pre-treatment patient stratification. Moreover, our radiomics-based score seems not to be associated with commonly studied clinical predictors such as anti-Scl-70 positivity or lung function, underlining a possible additive value to ‘traditional’ clinical parameters.References:[1]Schniering, J., et al. Resolving phenotypic and prognostic differences in interstitial lung disease related to systemic sclerosis by computed tomography-based radiomics. medRxiv [Preprint] doi:10.1101/2020.06.09.20124800 (2020).Disclosure of Interests:Chantal Meier: None declared, Malgorzata Maciukiewicz: None declared, Matthias Brunner: None declared, Janine Schniering: None declared, Hubert Gabrys: None declared, Anja Kühnis: None declared, Oliver Distler Speakers bureau: Speaker fee on Scleroderma and related complications: Bayer, Boehringer Ingelheim, Medscape, Novartis, Roche. Speaker fee on rheumatology topic other than Scleroderma: MSD, iQone, Novartis, Pfizer, Roche, Consultant of: Consultancy fee for Scleroderma and its complications: Abbvie, Acceleron Pharma, Amgen, AnaMar, Arxx Therapeutics, Bayer, Baecon Discovery, Boehringer, CSL Behring, ChemomAb, Corbus Pharmaceuticals, Horizon Pharmaceuticals, Galapagos NV, GSK, Glenmark Pharmaceuticals, Inventiva, Italfarmaco, iQvia, Kymera, Medac, Medscape, Mitsubishi Tanabe Pharma, MSD, Roche, Roivant Sciences, Sanofi, UCB. Consultancy fee for rheumatology topic other than Scleroderma: Abbvie, Amgen, Lilly, Pfizer, Grant/research support from: Research Grants to investigate the pathophysiology and potential treatment of Scleroderma and its complications: Kymera Therapeutics, Mitsubishi Tanabe, Thomas Frauenfelder: None declared, Stephanie Tanadini-Lang: None declared, Britta Maurer Speakers bureau: Speaker fees from Boehringer-Ingelheim, Grant/research support from: Grant/research support from AbbVie, Protagen, Novartis Biomedical Research, congress support from Pfizer, Roche, Actelion, mepha, and MSD
Inclusion of genetic variants in an ensemble of gradient boosting decision trees does not improve the prediction of citalopram treatment response
