ASSOCIATION OF A LOWER BODY-MASS INDEX WITH THE PRESENCE OF ILD IN SSc PATIENTS – A DERIVATION PREDICTION STUDY USING DECISION TREE-BASED ALGORITHMS

Background: Upper-gastrointestinal involvement (GI) is associated with more severe interstitial lung disease in patients with systemic sclerosis (SSc-ILD). However, there are many unexplored GI risk factors for the presence of SSc-ILD which could be potentially revealed by machine learning algorithms. Objectives: The aim of our study was to identify GI related risk factors for the presence of SSc-ILD using machine learning algorithms based on decision trees (DT). Methods: Data of the last follow-up visit from consecutive patients fulfilling the 2013 ACR/EULAR SSc classification criteria recorded in our local EUSTAR registry were used for this study. The study outcome was the presence of SSc-ILD on high-resolution computed tomography. Two sets of predictors were identified based on their potential association with GI. The first set contains the following variables available in the EUSTAR registry: e sophageal symptoms (dysphagia and reflux ), stomach symptoms (early satiety, vomiting ), intestinal symptoms (diarrhea, bloating and constipation ), malabsorption syndrome , body-mass-index (BMI ) and proton pump inhibitor therapy, calcium channel blocker therapy and immunosuppressive therapy . In the second set, we replaced the first three EUSTAR variables of the first set with the scales of the UCLA Gastrointestinal Tract Questionnaire 2.0 (UCLA-GIT). Of these two sets, the most important variable was selected using three different DT-based algorithms: (1) recursive partitioning and regression trees (RPART ) –which uses trees to build decision rules, (2) random forest (RF ) - an ensemble of DT built in parallel, and (3) gradient boosting machines (GBM ) - an ensemble of DT built sequentially. The selected variables were eventually integrated with established predictors for presence of SSc-ILD ( diffuse cutaneous subset , anti-Scl-70 positivity , male gender , forced vital capacity [FVC] and diffusion capacity of the lung for carbon monoxide-single breath [DLCO-SB] ) into final prediction models for SSc-ILD presence using RPART, RF and GBM respectively. Their performance was evaluated by C-statistics. The importance of the newly detected predictor was assessed by variable importance plots (VIPs). Results: We included in our study 334 patients. The median age was 61 [IQR: 50-69] years, 59 (17.7%) were males and 266 (79.6%) had limited cutaneous SSc. Median BMI was 23 [IQR: 21-26] kg/m2, 133 (39.8%) of the patients had SSc-ILD, median FVC% 93 [IQR: 81-105], DLCO 72.5 [56-84] and. Of the UCLA-GIT scales the highest score was for the distension/bloating with a value of 0.50 [IQR: 0-1.24]. Regarding medications, 167 (50%) patients were exposed to PPI, 39 (11.7%) to CCB and 105 (31.4%) to immunosuppressive therapy. The BMI was deemed by all three algorithms as the most important predictor of SSc-ILD among both sets of GI related variables ( Figure 1A -F). The final model, which included established risk factors for presence of ILD and the BMI, supported the importance of BMI in predicting the SSc-ILD. The VIPs obtained by GBM also ranked the BMI as the most important predictor. Figure 1. Tree-based algorithms revealing the importance of BMI for prediction of SSc-ILD. Panels A, B and C are variable importance plots (VIPs), which reveals the most important GI-predictor for occurrence of SSc-ILD in the EUSTAR set– the predictor with the highest relative importance is the most important predictor. Panels D, E and F are VIPs reveals the most important GI-predictor for occurrence of SSc-ILD in the UCLA-GIT set. A lower BMI was associated with presence of SSc-ILD (C-statistics for the RPART, RF and GBM models were 0.79, 0.70 and 0.76, respectively, corresponding to a fair accuracy). As expected, also a lower FVC, and DLCO-SB, and a positivity for Scl-70 ab were associated with presence of ILD. Conclusion: Lower BMI is a novel promising predictor for the presence of ILD, which should be confirmed in additional analyses. Disclosure of Interests: Alexandru Garaiman: None declared, Carina Mihai Speakers bureau: MEDtalks Switzerland, Mepha, Rucsandra Dobrota Consultant of: Actelion and Boehringer-Ingelheim, Grant/research support from: Articulum Fellowship, Pfizer, Actelion, Cosimo Bruni Speakers bureau: Eli-Lilly, Actelion, Boehringer-Ingelheim, Grant/research support from: Gruppo Italiano Lotta alla Sclerodermia (GILS), European,, Scleroderma Trials and Research Group (EUSTAR), Scleroderma Clinical Trials Consortium (SCTC), AbbVie, Muriel Elhai: None declared, Suzana Jordan: None declared, Lea Stamm: None declared, Anna-Maria Hoffmann-Vold Speakers bureau: Actelion, Boehringer Ingelheim, Jansen, Roche, Merck Sharp & Dohme, ARXX Therapeutics, Lilly and Medscape, Consultant of: Actelion, Boehringer Ingelheim, Jansen, Roche, Merck Sharp & Dohme, ARXX Therapeutics, Lilly and Medscape, Grant/research support from: Boehringer Ingelheim, Bayer, Oliver Distler Speakers bureau: Bayer, Boehringer Ingelheim, Medscape, Novartis, Roche, Pfizer, Roche, Sanofi, Consultant of: Abbvie, Acceleron, Alcimed, Amgen, AnaMar, Arxx, AstraZeneca, Baecon, Blade, Bayer, Boehringer Ingelheim, ChemomAb, Corbus, CSL Behring, Galapagos, Glenmark, GSK, Horizon, Inventiva, iQvia, Kymera, Lupin, Medac, Medscape, Miltenyi Biotec, Mitsubishi Tanabe, MSD, Prometheus Biosences, Roche, Roivant, Topadur and UBC, Lilly, Pfizer, Grant/research support from: Kymera, Mitsubishi Tanabe, Mike O. Becker Speakers bureau: Mepha, MSD, Novartis, GSK, Bayer and Vifor Citation: , volume 81, supplement 1, year 2022, page 734Session: Scleroderma, myositis and related syndromes (POSTERS only)